{"apiVersion":"v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/","scope":"Normalized factual records. Original source material, photographs, code, models and datasets retain their own licences."},"attribution":{"name":"humanoidversus.com","url":"https://humanoidversus.com/data","required":true,"text":"Data: humanoidversus.com. Link to the canonical record or /data, retain source references and verification dates, and indicate any changes."},"lastModified":"2026-10-13","data":[{"slug":"gemini-robotics-2","name":"Gemini Robotics 2","aliases":[],"org":{"name":"Google DeepMind","company":null},"type":"vla","released":"2026-07-30","summary":"Gemini Robotics 2 extends VLA control to humanoid walking, reaching and manipulation. Google DeepMind reports a shared checkpoint on Apollo 2 with different hands and on Franka Duo, with access limited to early-access partners.","openWeights":null,"license":null,"params":null,"baseModel":null,"inputs":["vision","language"],"outputs":["whole-body motor commands"],"embodiments":[{"robot":"apptronik-apollo-2","name":"Apptronik Apollo 2","note":"Demonstrated with SharpaWave and Inspire hands."},{"robot":null,"name":"Franka Duo","note":"Demonstrated with a Robotiq gripper."}],"trainingData":[],"links":{"paper":null,"code":null,"weights":null,"project":"https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/"},"humanoid":"demonstrated","sources":[{"id":"s1","url":"https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/","title":"Gemini Robotics 2 brings whole body intelligence to robots","publisher":"Google DeepMind","date":"2026-07-30","type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/gemini-robotics-2"},{"slug":"gemini-robotics-er-2","name":"Gemini Robotics ER 2","aliases":[],"org":{"name":"Google DeepMind","company":null},"type":"foundation-model","released":"2026-07-30","summary":"Gemini Robotics ER 2 is the high-level reasoning agent in the Gemini Robotics 2 system. It plans task sequences, coordinates VLA execution and robot collaboration, and tracks progress; Google offers it through AI Studio.","openWeights":null,"license":null,"params":null,"baseModel":"Gemini 3.5 Flash","inputs":["vision","language"],"outputs":["task plans","VLA instructions","progress estimates"],"embodiments":[],"trainingData":[{"dataset":null,"name":"Gemini 3.5 training datasets and embodied-reasoning datasets"}],"links":{"paper":null,"code":null,"weights":null,"project":"https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/","title":"Gemini Robotics 2 brings whole body intelligence to robots","publisher":"Google DeepMind","date":"2026-07-30","type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://deepmind.google/models/model-cards/gemini-robotics-er-2/","title":"Gemini Robotics ER 2 model card","publisher":"Google DeepMind","date":"2026-07-30","type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/gemini-robotics-er-2"},{"slug":"gemini-robotics-on-device-2","name":"Gemini Robotics On-Device 2","aliases":[],"org":{"name":"Google DeepMind","company":null},"type":"vla","released":"2026-07-30","summary":"Gemini Robotics On-Device 2 is a local VLA with multi-embodiment motion-transfer training. The announcement demonstrates adaptation to Dexmate, SO101 and Trossen platforms and limits availability to early-access partners.","openWeights":null,"license":null,"params":null,"baseModel":null,"inputs":["vision","language"],"outputs":["robot actions"],"embodiments":[{"robot":null,"name":"Dexmate platform","note":"Exact robot generation is not specified in the announcement."},{"robot":null,"name":"SO101","note":null},{"robot":null,"name":"Trossen platform","note":"Exact configuration is not specified in the announcement."}],"trainingData":[{"dataset":null,"name":"Images, text, robot sensor data and action data"}],"links":{"paper":null,"code":null,"weights":null,"project":"https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/","title":"Gemini Robotics 2 brings whole body intelligence to robots","publisher":"Google DeepMind","date":"2026-07-30","type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://deepmind.google/models/model-cards/gemini-robotics-on-device-2/","title":"Gemini Robotics On-Device 2 model card","publisher":"Google DeepMind","date":"2026-07-30","type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/gemini-robotics-on-device-2"},{"slug":"lingbot-vla-2-0","name":"LingBot-VLA 2.0","aliases":["LingBot-VLA-V2","LingBot-VLA-v2-6B"],"org":{"name":"Robbyant","company":"robbyant"},"type":"vla","released":"2026-07-08","summary":"LingBot-VLA 2.0 expands the training mixture, action representation and predictive-dynamics objectives of LingBot-VLA. Its native-depth checkpoint uses Qwen3-VL, and the repository reports real manipulation results on Galaxea R1 Pro.","openWeights":true,"license":"Apache-2.0","params":6000000000,"baseModel":"Qwen3-VL-4B-Instruct","inputs":["vision","language","depth"],"outputs":["robot actions"],"embodiments":[{"robot":"galaxea-r1","name":"Galaxea R1 Pro","note":"Real-robot evaluation table in the official repository; the catalog groups the R1 variants."}],"trainingData":[{"dataset":null,"name":"Robbyant robot-trajectory and egocentric human-video mixture"}],"links":{"paper":"https://arxiv.org/abs/2607.06403","code":"https://github.com/Robbyant/lingbot-vla-v2","weights":"https://huggingface.co/robbyant/lingbot-vla-v2-6b","project":null},"humanoid":"demonstrated","sources":[{"id":"s1","url":"https://raw.githubusercontent.com/Robbyant/lingbot-vla-v2/main/README.md","title":"LingBot-VLA 2.0 official README: release, training and real-robot evaluations","publisher":"Robbyant","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://huggingface.co/robbyant/lingbot-vla-v2-6b","title":"LingBot-VLA 2.0 6B official model card","publisher":"Robbyant","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://arxiv.org/abs/2607.06403","title":"From Foundation to Application: Improving VLA Models in Practice","publisher":"Robbyant","date":"2026-07-07","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/lingbot-vla-2-0"},{"slug":"galaxea-g0-5","name":"Galaxea G0.5","aliases":["G05"],"org":{"name":"Galaxea AI","company":"galaxea"},"type":"vla","released":"2026-06-01","summary":"Galaxea G0.5 predicts reasoning and action tokens in one autoregressive stream. Its public repository supplies pretrained checkpoints and real-robot deployment paths for R1 Lite, R1 Pro, DROID and SO100/101.","openWeights":true,"license":"G0.5 COMMUNITY LICENSE AGREEMENT (NON-COMMERCIAL + LIMITED PATENT LICENSE)","params":null,"baseModel":"Qwen3.5 2B VLM","inputs":["vision","language","proprioception","embodiment identifier"],"outputs":["reasoning tokens","action tokens decoded to continuous controls"],"embodiments":[{"robot":"galaxea-r1","name":"Galaxea R1 Lite / R1 Pro","note":"The catalog groups these variants; separate deployment entrypoints are published."},{"robot":null,"name":"DROID Franka setup","note":null},{"robot":null,"name":"SO100 / SO101","note":null}],"trainingData":[{"dataset":null,"name":"Fourteen-embodiment foundation robot mixture with web and embodied VQA data"},{"dataset":"droid","name":"DROID (subsequent DROID post-training; excluded from foundation pretraining)"},{"dataset":"libero-demos","name":"LIBERO (post-trained LIBERO checkpoint)"},{"dataset":"robotwin-2","name":"RoboTwin 2.0 (post-trained evaluation checkpoint)"},{"dataset":"behavior-1k-2025-challenge-demos","name":"BEHAVIOR 2025 challenge demonstrations (single-policy post-training)"}],"links":{"paper":"https://arxiv.org/abs/2608.11739","code":"https://github.com/OpenGalaxea/GalaxeaVLA","weights":"https://huggingface.co/OpenGalaxea/G05","project":"https://opengalaxea.github.io/G05/"},"humanoid":"demonstrated","sources":[{"id":"s1","url":"https://raw.githubusercontent.com/OpenGalaxea/GalaxeaVLA/main/README.md","title":"G0.5 official README: announcement, method and deployment","publisher":"Galaxea AI","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://github.com/OpenGalaxea/GalaxeaVLA/blob/main/LICENSE-G0.5","title":"G0.5 Community License Agreement","publisher":"Galaxea AI","date":"2026-06-16","type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://arxiv.org/html/2608.11739v1","title":"G0.5 paper: model design, data mixture and R1 experiments","publisher":"Galaxea AI","date":"2026-08-12","type":"paper","accessed":"2026-10-13"},{"id":"s4","url":"https://huggingface.co/OpenGalaxea/G05","title":"Galaxea G0.5 model card, inputs, checkpoint variants and licensing","publisher":"Galaxea AI","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/galaxea-g0-5"},{"slug":"wall-oss-0-5","name":"Wall-OSS-0.5","aliases":[],"org":{"name":"X Square Robot","company":"x-square-robot"},"type":"vla","released":"2026-05-29","summary":"Wall-OSS-0.5 co-trains action and multimodal prediction using Qwen2.5-VL-3B-Instruct. The report labels the model 4B but describes the full architecture as over 4B, so an exact total is not assigned; independently stated checkpoint license terms were not established.","openWeights":true,"license":null,"params":null,"baseModel":"Qwen2.5-VL-3B-Instruct","inputs":["vision","language","proprioception"],"outputs":["continuous robot action chunks"],"embodiments":[],"trainingData":[{"dataset":"robomind","name":"RoboMIND v1"},{"dataset":"robomind-2-0","name":"RoboMIND v2.0"},{"dataset":"agibot-world","name":"AgiBotWorld Beta"},{"dataset":"galaxea-open-world","name":"Galaxea Open-World"},{"dataset":"droid","name":"DROID"},{"dataset":"bridgedata-v2","name":"BridgeData V2"},{"dataset":null,"name":"Self-collected data, other open robot trajectories and grounded multimodal data"}],"links":{"paper":"https://arxiv.org/abs/2605.30877","code":"https://github.com/X-Square-Robot/wall-x","weights":"https://huggingface.co/x-square-robot/wall-oss-0.5","project":null},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2605.30877","title":"Wall-OSS-0.5 Technical Report","publisher":"X Square Robot","date":"2026-05-29","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://huggingface.co/x-square-robot/wall-oss-0.5","title":"Wall-OSS-0.5 official model card","publisher":"X Square Robot","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://arxiv.org/html/2605.30877v2","title":"Wall-OSS-0.5 report: parameter scope, initialization and training mixture","publisher":"X Square Robot","date":"2026-06-01","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/wall-oss-0-5"},{"slug":"gr00t-n1-7","name":"Isaac GR00T N1.7","aliases":["GR00T N1.7"],"org":{"name":"NVIDIA","company":null},"type":"vla","released":"2026-04-18","summary":"GR00T N1.7 is a cross-embodiment VLA with a Cosmos-Reason2-2B backbone and a published 3B checkpoint. The checkpoint card specifies NVIDIA Open Model License Agreement terms, while the current repository describes Apache-2.0 licensing and general availability.","openWeights":true,"license":"NVIDIA Open Model License Agreement","params":3000000000,"baseModel":"Cosmos-Reason2-2B","inputs":["vision","language","proprioception"],"outputs":["continuous action chunks"],"embodiments":[{"robot":"unitree-g1","name":"Unitree G1","note":"NVIDIA's May 2026 workflow documents GR00T N1.7 fine-tuning and deployment with SONIC whole-body control."}],"trainingData":[{"dataset":null,"name":"EgoScale human video and heterogeneous robot demonstrations"},{"dataset":"droid","name":"DROID (post-trained DROID variant)"},{"dataset":"libero-demos","name":"LIBERO (post-trained LIBERO variant)"},{"dataset":"bridgedata-v2","name":"BridgeData V2 (post-trained SimplerEnv Bridge variant)"}],"links":{"paper":null,"code":"https://github.com/NVIDIA/Isaac-GR00T","weights":"https://huggingface.co/nvidia/GR00T-N1.7-3B","project":null},"humanoid":"native","sources":[{"id":"s1","url":"https://huggingface.co/nvidia/GR00T-N1.7-3B","title":"GR00T-N1.7-3B model card","publisher":"NVIDIA","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://github.com/NVIDIA/Isaac-GR00T/releases/tag/n1.7-release","title":"Isaac GR00T N1.7 early-access repository release","publisher":"NVIDIA","date":"2026-04-18","type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://raw.githubusercontent.com/NVlabs/GR00T-WholeBodyControl/main/README.md","title":"GR00T Whole-Body Control: end-to-end GR00T N1.7 workflow on G1","publisher":"NVIDIA","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://github.com/NVIDIA/Isaac-GR00T","title":"Isaac GR00T official repository: model versions and code licensing","publisher":"NVIDIA","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/gr00t-n1-7"},{"slug":"pi0-7","name":"π0.7","aliases":["pi0.7"],"org":{"name":"Physical Intelligence","company":null},"type":"vla","released":"2026-04-16","summary":"π0.7 is a generalist VLA steered by language, episode metadata, control mode and optional subgoal images, with a video-history encoder. The paper reports approximately 5B total parameters and evaluations on mobile, static bimanual and UR5e platforms.","openWeights":null,"license":null,"params":5000000000,"baseModel":"Gemma 3 4B","inputs":["vision","language","proprioception","subgoal images","episode metadata","control mode"],"outputs":["continuous robot action chunks"],"embodiments":[{"robot":null,"name":"Bimanual mobile manipulator","note":"Two 6-DoF arms; exact platform model is not identified."},{"robot":null,"name":"BiPi","note":"Static bimanual system with lightweight 6-DoF arms."},{"robot":null,"name":"Bimanual UR5e","note":"Robotiq parallel-jaw grippers; cross-embodiment evaluation."},{"robot":null,"name":"Single-arm BiPi-derived setup","note":"Used in generalization and language-following experiments."}],"trainingData":[{"dataset":null,"name":"Multi-robot demonstrations, autonomous evaluations, human interventions, open robot datasets, egocentric human videos and web data"}],"links":{"paper":"https://arxiv.org/abs/2604.15483","code":null,"weights":null,"project":null},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2604.15483","title":"π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities","publisher":"Physical Intelligence","date":"2026-04-16","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://arxiv.org/html/2604.15483v2","title":"π0.7 paper: architecture, parameter scope, training and embodiments","publisher":"Physical Intelligence","date":"2026-04-24","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/pi0-7"},{"slug":"gemini-robotics-er-1-6","name":"Gemini Robotics-ER 1.6","aliases":[],"org":{"name":"Google DeepMind","company":null},"type":"foundation-model","released":"2026-04-14","summary":"Gemini Robotics-ER 1.6 updates embodied reasoning with multi-view understanding, success detection and instrument reading. Google DeepMind describes facility-inspection work with Boston Dynamics Spot and makes the model available through the Gemini API.","openWeights":null,"license":null,"params":null,"baseModel":null,"inputs":["vision","language"],"outputs":["task plans","tool calls","spatial predictions","instrument readings"],"embodiments":[{"robot":null,"name":"Boston Dynamics Spot","note":"Quadruped facility-inspection use case; outside the humanoid robot collection."}],"trainingData":[],"links":{"paper":null,"code":null,"weights":null,"project":"https://deepmind.google/blog/gemini-robotics-er-1-6/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://deepmind.google/blog/gemini-robotics-er-1-6/","title":"Gemini Robotics-ER 1.6: enhanced embodied reasoning","publisher":"Google DeepMind","date":"2026-04-14","type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/gemini-robotics-er-1-6"},{"slug":"dreamzero","name":"DreamZero","aliases":["DreamZero-DROID","DreamZero-AgiBot"],"org":{"name":"NVIDIA GEAR Lab","company":null},"type":"world-model","released":"2026-02-17","summary":"DreamZero jointly predicts future video and robot actions with a video-diffusion backbone; its paper reports a 14B model. DreamZero-DROID weights use CC BY-NC 4.0, DreamZero-AgiBot weights are tagged Apache-2.0, and repository code uses Apache-2.0.","openWeights":true,"license":"CC BY-NC 4.0","params":14000000000,"baseModel":"Wan2.1-I2V-14B-480P","inputs":["vision","language","proprioception"],"outputs":["predicted video","robot action chunks"],"embodiments":[{"robot":"agibot-genie-g1","name":"AgiBot G1","note":"AgiBot checkpoint and real-robot experiments; separate from the DROID checkpoint."},{"robot":null,"name":"Franka DROID setup","note":"DreamZero-DROID checkpoint and real-robot evaluation."},{"robot":null,"name":"YAM","note":"Few-shot embodiment adaptation from the AgiBot model."}],"trainingData":[{"dataset":"droid","name":"DROID 1.0.1, filtered LeRobot conversion for DreamZero-DROID"},{"dataset":null,"name":"Self-collected AgiBot G1 teleoperation corpus for DreamZero-AgiBot"},{"dataset":null,"name":"YAM play data and video-only human/robot demonstrations for transfer experiments"}],"links":{"paper":"https://arxiv.org/abs/2602.15922","code":"https://github.com/dreamzero0/dreamzero","weights":"https://huggingface.co/GEAR-Dreams/DreamZero-DROID","project":"https://dreamzero0.github.io/"},"humanoid":"demonstrated","sources":[{"id":"s1","url":"https://arxiv.org/abs/2602.15922","title":"World Action Models are Zero-shot Policies","publisher":"NVIDIA GEAR Lab","date":"2026-02-17","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://arxiv.org/html/2602.15922v1","title":"DreamZero paper: model, robot experiments and separate embodiment training","publisher":"NVIDIA GEAR Lab","date":"2026-02-17","type":"paper","accessed":"2026-10-13"},{"id":"s3","url":"https://raw.githubusercontent.com/dreamzero0/dreamzero/main/README.md","title":"DreamZero official README: Wan backbone, checkpoint releases and code license","publisher":"NVIDIA GEAR Lab","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://huggingface.co/GEAR-Dreams/DreamZero-DROID","title":"DreamZero-DROID model card: 14B model and CC BY-NC 4.0 tag","publisher":"NVIDIA GEAR Lab","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s5","url":"https://huggingface.co/GEAR-Dreams/DreamZero-AgiBot","title":"DreamZero-AgiBot checkpoint and Apache-2.0 tag","publisher":"NVIDIA GEAR Lab","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/dreamzero"},{"slug":"galaxea-g0-tiny","name":"Galaxea G0Tiny","aliases":["G0Tiny","G0Tiny_250M_base"],"org":{"name":"Galaxea AI","company":"galaxea"},"type":"vla","released":"2026-02-12","summary":"G0Tiny is a 250M VLA for local inference on Galaxea R1 Pro Orin, pretrained on R1 Pro VR teleoperation data. The pinned release README assigns the G0 PLUS Community License to the release material; the shared Hub repository still displays a CC BY-NC-SA 4.0 tag.","openWeights":true,"license":"G0 PLUS COMMUNITY LICENSE AGREEMENT (NON-COMMERCIAL + LIMITED PATENT LICENSE)","params":250000000,"baseModel":"SmolVLM2-500M-Video-Instruct","inputs":["vision","language","proprioception"],"outputs":["continuous robot action chunks"],"embodiments":[{"robot":"galaxea-r1","name":"Galaxea R1 Pro Orin","note":"R1-family catalog link; G0Tiny's Fold Towels and Handover Gift demos use the R1 Pro Orin variant."}],"trainingData":[{"dataset":null,"name":"R1 Pro VR teleoperation demonstrations"}],"links":{"paper":null,"code":"https://github.com/OpenGalaxea/GalaxeaVLA/tree/13a16a9049aee8f1d799b56fccc0c5832a75fc2f","weights":"https://huggingface.co/OpenGalaxea/G0-VLA/tree/main/G0Tiny_260120","project":null},"humanoid":"demonstrated","sources":[{"id":"s1","url":"https://raw.githubusercontent.com/OpenGalaxea/GalaxeaVLA/13a16a9049aee8f1d799b56fccc0c5832a75fc2f/README.md","title":"Pinned G0Plus/G0Tiny README: release, model size, backbone, R1 Pro demos and license scope","publisher":"Galaxea AI","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://huggingface.co/OpenGalaxea/G0-VLA/tree/main/G0Tiny_260120","title":"G0Tiny checkpoint directory with inherited shared-repository license metadata","publisher":"Galaxea AI","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://raw.githubusercontent.com/OpenGalaxea/GalaxeaVLA/main/licenses/LICENSE-G0Plus","title":"G0 PLUS Community License Agreement: scope includes code and weights","publisher":"Galaxea AI","date":"2026-01-04","type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://huggingface.co/OpenGalaxea/G0-VLA/blob/main/G0Tiny_260120/config.yaml","title":"G0Tiny checkpoint configuration: backbone and observation fields","publisher":"Galaxea AI","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/galaxea-g0-tiny"},{"slug":"unifolm-vla-0","name":"UnifoLM-VLA-0","aliases":[],"org":{"name":"Unitree Robotics","company":"unitree"},"type":"vla","released":"2026-01-29","summary":"UnifoLM-VLA-0 is Unitree's language-conditioned model for humanoid manipulation. Its release includes continued visual-language pretraining, robot-action training, and a checkpoint fine-tuned on Unitree's open manipulation datasets.","openWeights":true,"license":"CC BY-NC-SA 4.0","params":null,"baseModel":"UnifoLM-VLM-Base","inputs":["vision","language","proprioception"],"outputs":["robot actions"],"embodiments":[{"robot":"unitree-g1","name":"Unitree G1","note":"The official README identifies G1 block-stacking and bag-insertion examples."}],"trainingData":[{"dataset":null,"name":"Unitree open G1 manipulation datasets"},{"dataset":null,"name":"General image-text VQA and open robot datasets"}],"links":{"paper":null,"code":"https://github.com/unitreerobotics/unifolm-vla","weights":"https://huggingface.co/unitreerobotics/UnifoLM-VLA-Base","project":null},"humanoid":"native","sources":[{"id":"s1","url":"https://raw.githubusercontent.com/unitreerobotics/unifolm-vla/main/README.md","title":"UnifoLM-VLA-0 official README and release date","publisher":"Unitree Robotics","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://huggingface.co/unitreerobotics/UnifoLM-VLA-Base","title":"UnifoLM-VLA-Base model card","publisher":"Unitree Robotics","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/unifolm-vla-0"},{"slug":"helix-02","name":"Helix 02","aliases":[],"org":{"name":"Figure AI","company":"figure-ai"},"type":"vla","released":"2026-01-27","summary":"Helix 02 extends Figure's VLA to coordinated whole-body locomotion and manipulation. It adds a learned System 0 controller and incorporates Figure 03 palm cameras and fingertip tactile sensing alongside vision and full-body state.","openWeights":null,"license":null,"params":null,"baseModel":null,"inputs":["vision","language","proprioception","touch"],"outputs":["whole-body joint commands"],"embodiments":[{"robot":"figure-03","name":"Figure 03","note":"The announcement explicitly identifies Figure 03 sensing hardware."}],"trainingData":[{"dataset":null,"name":"Retargeted human motion and simulation for System 0"}],"links":{"paper":null,"code":null,"weights":null,"project":"https://www.figure.ai/news/helix-02"},"humanoid":"native","sources":[{"id":"s1","url":"https://www.figure.ai/news/helix-02","title":"Introducing Helix 02: Full-Body Autonomy","publisher":"Figure AI","date":"2026-01-27","type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/helix-02"},{"slug":"lingbot-vla","name":"LingBot-VLA","aliases":["LingBot-VLA-4B"],"org":{"name":"Robbyant","company":"robbyant"},"type":"vla","released":"2026-01-27","summary":"LingBot-VLA is a manipulation foundation model pretrained on real-world dual-arm robot trajectories. Robbyant publishes depth-free and depth-distilled checkpoints, post-training code and evaluations on AgiBot G1 and Galaxea R1 Pro.","openWeights":true,"license":"Apache-2.0","params":4000000000,"baseModel":"Qwen2.5-VL-3B-Instruct","inputs":["vision","language"],"outputs":["robot actions"],"embodiments":[{"robot":"agibot-genie-g1","name":"AgiBot G1","note":"Real-robot evaluation table in the official repository."},{"robot":"galaxea-r1","name":"Galaxea R1 Pro","note":"R1 Pro evaluation; the catalog groups the R1 variants."}],"trainingData":[{"dataset":null,"name":"Robbyant multi-configuration dual-arm robot trajectories"}],"links":{"paper":"https://arxiv.org/abs/2601.18692","code":"https://github.com/Robbyant/lingbot-vla","weights":"https://huggingface.co/robbyant/lingbot-vla-4b","project":null},"humanoid":"demonstrated","sources":[{"id":"s1","url":"https://raw.githubusercontent.com/Robbyant/lingbot-vla/main/README.md","title":"LingBot-VLA official README: release, backbone and real-robot evaluations","publisher":"Robbyant","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://huggingface.co/robbyant/lingbot-vla-4b","title":"LingBot-VLA-4B official model card","publisher":"Robbyant","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://arxiv.org/abs/2601.18692","title":"A Pragmatic VLA Foundation Model","publisher":"Robbyant","date":"2026-01-26","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/lingbot-vla"},{"slug":"gr00t-n1-6","name":"Isaac GR00T N1.6","aliases":["GR00T N1.6"],"org":{"name":"NVIDIA","company":null},"type":"vla","released":"2025-12-15","summary":"GR00T N1.6 uses an internal Cosmos-2B VLM and reports post-training evaluations on YAM arms, AgiBot Genie-1 and Unitree G1. Its checkpoint is labeled 3B; weights use NVIDIA OneWay Noncommercial License terms and code uses Apache-2.0.","openWeights":true,"license":"NVIDIA OneWay Noncommercial License","params":3000000000,"baseModel":"NVIDIA Cosmos-2B VLM (internal variant)","inputs":["vision","language","proprioception"],"outputs":["continuous action chunks"],"embodiments":[{"robot":"unitree-g1","name":"Unitree G1","note":"Real-robot evaluation and whole-body locomotion/manipulation training data."},{"robot":"agibot-genie-g1","name":"AgiBot Genie-1","note":"Named Genie-1 on NVIDIA's research page."},{"robot":null,"name":"Bimanual YAM arms","note":"Real-robot evaluation."}],"trainingData":[{"dataset":null,"name":"Bimanual YAM and AgiBot Genie-1 teleoperation"},{"dataset":null,"name":"Simulated Galaxea R1 Pro on BEHAVIOR"},{"dataset":null,"name":"Unitree G1 whole-body locomotion and manipulation demonstrations"}],"links":{"paper":null,"code":"https://github.com/NVIDIA/Isaac-GR00T","weights":"https://huggingface.co/nvidia/GR00T-N1.6-3B","project":"https://research.nvidia.com/labs/gear/gr00t-n1_6/"},"humanoid":"native","sources":[{"id":"s1","url":"https://huggingface.co/nvidia/GR00T-N1.6-3B","title":"GR00T-N1.6-3B model card","publisher":"NVIDIA","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://research.nvidia.com/labs/gear/gr00t-n1_6/","title":"GR00T N1.6: An Improved Open Foundation Model for Generalist Humanoid Robots","publisher":"NVIDIA","date":"2025-12-15","type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://github.com/NVIDIA/Isaac-GR00T","title":"Isaac GR00T official repository: model versions and code licensing","publisher":"NVIDIA","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://developer.download.nvidia.com/licenses/NVIDIA-OneWay-Noncommercial-License-22Mar2022.pdf","title":"NVIDIA OneWay Noncommercial License (22 March 2022)","publisher":"NVIDIA","date":"2022-03-22","type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/gr00t-n1-6"},{"slug":"pi-star-0-6","name":"π*0.6","aliases":["pi-star-0.6","π⁎0.6"],"org":{"name":"Physical Intelligence","company":null},"type":"vla","released":"2025-11-18","summary":"π*0.6 applies RECAP reinforcement learning to a generalist VLA using demonstrations, autonomous rollouts and teleoperated corrections. The paper evaluates iterative improvement on a static bimanual robot for laundry folding, box assembly and espresso preparation.","openWeights":null,"license":null,"params":null,"baseModel":"Gemma 3 4B (π0.6 architecture)","inputs":["vision","language","proprioception","advantage condition"],"outputs":["continuous joint-position and gripper action chunks","subtask text"],"embodiments":[{"robot":null,"name":"Static bimanual system with two 6-DoF arms","note":"Parallel-jaw grippers; exact commercial arm model is not identified."}],"trainingData":[{"dataset":null,"name":"Multi-robot demonstrations, web vision-language data, autonomous rollouts and expert interventions"}],"links":{"paper":"https://arxiv.org/abs/2511.14759","code":null,"weights":null,"project":null},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2511.14759","title":"π*0.6: a VLA That Learns From Experience","publisher":"Physical Intelligence","date":"2025-11-18","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://arxiv.org/html/2511.14759v2","title":"π*0.6 paper: Gemma 3 backbone, RECAP training and robot setup","publisher":"Physical Intelligence","date":"2025-11-19","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/pi-star-0-6"},{"slug":"sonic","name":"SONIC","aliases":["GEAR-SONIC"],"org":{"name":"NVIDIA","company":null},"type":"whole-body-controller","released":"2025-11-11","summary":"SONIC learns general humanoid motion tracking and supplies Unitree G1 controller variants for planning, teleoperation and VLA execution. Model weights use NVIDIA Open Model License terms; source code uses Apache-2.0.","openWeights":true,"license":"NVIDIA Open Model License","params":null,"baseModel":null,"inputs":["robot state","reference motion","latent motion tokens"],"outputs":["whole-body motor commands"],"embodiments":[{"robot":"unitree-g1","name":"Unitree G1","note":"Published controller checkpoints and real-robot deployment stack; variants must use matching encoder, decoder and observation configuration."}],"trainingData":[{"dataset":null,"name":"Human motion-capture training corpus"}],"links":{"paper":"https://arxiv.org/abs/2511.07820","code":"https://github.com/NVlabs/GR00T-WholeBodyControl","weights":"https://huggingface.co/nvidia/GEAR-SONIC","project":"https://nvlabs.github.io/GEAR-SONIC/"},"humanoid":"native","sources":[{"id":"s1","url":"https://arxiv.org/abs/2511.07820","title":"SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control","publisher":"NVIDIA","date":"2025-11-11","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://huggingface.co/nvidia/GEAR-SONIC","title":"GEAR-SONIC official model card and checkpoint variants","publisher":"NVIDIA","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://raw.githubusercontent.com/NVlabs/GR00T-WholeBodyControl/main/README.md","title":"GR00T Whole-Body Control: SONIC checkpoints, deployment and licensing","publisher":"NVIDIA","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/sonic"},{"slug":"gemini-robotics-1-5","name":"Gemini Robotics 1.5","aliases":[],"org":{"name":"Google DeepMind","company":null},"type":"vla","released":"2025-09-25","summary":"Gemini Robotics 1.5 combines motor action generation with natural-language reasoning about actions. Its motion-transfer training supports demonstrations across ALOHA 2, Apptronik Apollo and bimanual Franka embodiments.","openWeights":null,"license":null,"params":null,"baseModel":"Gemini family","inputs":["vision","language"],"outputs":["robot motor commands","natural-language reasoning"],"embodiments":[{"robot":null,"name":"ALOHA 2","note":null},{"robot":null,"name":"Bimanual Franka platform","note":null},{"robot":"apptronik-apollo","name":"Apptronik Apollo","note":"Humanoid adaptation demonstrated in the announcement."}],"trainingData":[],"links":{"paper":"https://arxiv.org/abs/2510.03342","code":null,"weights":null,"project":"https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/"},"humanoid":"demonstrated","sources":[{"id":"s1","url":"https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/","title":"Gemini Robotics 1.5 brings AI agents into the physical world","publisher":"Google DeepMind","date":"2025-09-25","type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://arxiv.org/abs/2510.03342","title":"Gemini Robotics 1.5 technical report","publisher":"Google DeepMind","date":"2025-10-02","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/gemini-robotics-1-5"},{"slug":"gemini-robotics-er-1-5","name":"Gemini Robotics-ER 1.5","aliases":[],"org":{"name":"Google DeepMind","company":null},"type":"foundation-model","released":"2025-09-25","summary":"Gemini Robotics-ER 1.5 supplies high-level planning, spatial reasoning and tool use for robot agents. The system passes language instructions to a VLA and estimates task progress and completion; developer access is through the Gemini API.","openWeights":null,"license":null,"params":null,"baseModel":"Gemini family","inputs":["vision","language"],"outputs":["task plans","tool calls","spatial predictions","progress estimates"],"embodiments":[],"trainingData":[],"links":{"paper":"https://arxiv.org/abs/2510.03342","code":null,"weights":null,"project":"https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/","title":"Gemini Robotics 1.5 brings AI agents into the physical world","publisher":"Google DeepMind","date":"2025-09-25","type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://arxiv.org/abs/2510.03342","title":"Gemini Robotics 1.5 technical report","publisher":"Google DeepMind","date":"2025-10-02","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/gemini-robotics-er-1-5"},{"slug":"go-1","name":"GO-1","aliases":[],"org":{"name":"AgiBot World team / OpenDriveLab","company":"agibot"},"type":"vla","released":"2025-09-19","summary":"GO-1 is the AgiBot World team's VLA pretrained on its robot manipulation dataset. The public model uses an InternVL 2.5 backbone; the open-release study examines training and evaluation practices using AgiBot G1.","openWeights":true,"license":"CC BY-NC-SA 4.0","params":null,"baseModel":"InternVL 2.5-2B","inputs":["vision","language"],"outputs":["robot actions"],"embodiments":[{"robot":"agibot-genie-g1","name":"AgiBot G1","note":"The repository identifies the G1 as the Genie Studio platform for GO-1."}],"trainingData":[{"dataset":"agibot-world","name":"AgiBot World"}],"links":{"paper":"https://arxiv.org/abs/2503.06669","code":"https://github.com/OpenDriveLab/AgiBot-World","weights":"https://huggingface.co/agibot-world/GO-1","project":"https://opendrivelab.com/OpenGO1/"},"humanoid":"demonstrated","sources":[{"id":"s1","url":"https://huggingface.co/agibot-world/GO-1","title":"GO-1 official model card","publisher":"AgiBot World team","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://raw.githubusercontent.com/OpenDriveLab/AgiBot-World/main/README.md","title":"AgiBot World repository: GO-1 release and Genie Studio","publisher":"AgiBot World team / OpenDriveLab","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://opendrivelab.com/OpenGO1/","title":"Open-sourcing GO-1: building VLA systems at scale","publisher":"OpenDriveLab","date":"2025-09-19","type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/go-1"},{"slug":"unifolm-wma-0","name":"UnifoLM-WMA-0","aliases":[],"org":{"name":"Unitree Robotics","company":"unitree"},"type":"world-model","released":"2025-09-15","summary":"UnifoLM-WMA-0 predicts robot interactions and actions, with G1 camera-packing deployment examples. Its model-card body specifies CC BY-NC-SA 4.0 for the model despite an Apache-2.0 metadata tag.","openWeights":true,"license":"CC BY-NC-SA 4.0","params":null,"baseModel":null,"inputs":["vision","language","proprioception","actions (simulation mode)"],"outputs":["predicted interaction videos","robot action chunks"],"embodiments":[{"robot":"unitree-g1","name":"Unitree G1","note":"The official README includes a G1 camera-packing demonstration and deployment command."}],"trainingData":[{"dataset":"open-x-embodiment","name":"Open X-Embodiment"},{"dataset":null,"name":"Unitree G1 manipulation datasets"}],"links":{"paper":null,"code":"https://github.com/unitreerobotics/unifolm-world-model-action","weights":"https://huggingface.co/unitreerobotics/UnifoLM-WMA-0-Base","project":null},"humanoid":"demonstrated","sources":[{"id":"s1","url":"https://raw.githubusercontent.com/unitreerobotics/unifolm-world-model-action/main/README.md","title":"UnifoLM-WMA-0 official README","publisher":"Unitree Robotics","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://huggingface.co/unitreerobotics/UnifoLM-WMA-0-Base","title":"UnifoLM-WMA-0-Base model card","publisher":"Unitree Robotics","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/unifolm-wma-0"},{"slug":"wall-oss","name":"WALL-OSS","aliases":[],"org":{"name":"X Square Robot","company":"x-square-robot"},"type":"vla","released":"2025-09-15","summary":"WALL-OSS couples a Qwen2.5-VL-3B backbone to action generation and multimodal reasoning. Wall-X code uses Apache-2.0, but separately stated checkpoint license terms were not established.","openWeights":true,"license":null,"params":null,"baseModel":"Qwen2.5-VL-3B","inputs":["vision","language","proprioception"],"outputs":["robot action chunks"],"embodiments":[],"trainingData":[{"dataset":"agibot-world","name":"AgiBot World"},{"dataset":"droid","name":"DROID"},{"dataset":"rh20t","name":"RH20T"},{"dataset":"bridgedata-v2","name":"BridgeData V2"},{"dataset":null,"name":"Self-collected robot trajectories, other open robot datasets and multimodal VQA"}],"links":{"paper":"https://arxiv.org/abs/2509.11766","code":"https://github.com/X-Square-Robot/wall-x","weights":"https://huggingface.co/x-square-robot/wall-oss-flow","project":null},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2509.11766","title":"Igniting VLMs toward the Embodied Space","publisher":"X Square Robot","date":"2025-09-15","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://huggingface.co/x-square-robot/wall-oss-flow/raw/main/README.md","title":"WALL-OSS flow model card and inference example","publisher":"X Square Robot","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://github.com/X-Square-Robot/wall-x","title":"Wall-X official repository and model release branches","publisher":"X Square Robot","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://arxiv.org/html/2509.11766v1","title":"WALL-OSS paper: backbone and named training mixture","publisher":"X Square Robot","date":"2025-09-15","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/wall-oss"},{"slug":"galaxea-g0","name":"Galaxea G0","aliases":["G0-VLA"],"org":{"name":"Galaxea AI","company":"galaxea"},"type":"vla","released":"2025-08-30","summary":"Galaxea G0 pairs a visual-language task planner with a VLA and is evaluated on R1 Lite mobile manipulation. The original G0-VLA model card specifies CC BY-NC-SA 4.0 for weights; legacy repository code predating January 4, 2026 uses Apache-2.0.","openWeights":true,"license":"CC BY-NC-SA 4.0","params":null,"baseModel":null,"inputs":["vision","language"],"outputs":["task instructions","continuous robot actions"],"embodiments":[{"robot":"galaxea-r1","name":"Galaxea R1 Lite","note":"The catalog groups R1 variants; the original G0 project explicitly uses R1 Lite."}],"trainingData":[{"dataset":"galaxea-open-world","name":"Galaxea Open-World Dataset"}],"links":{"paper":"https://arxiv.org/abs/2509.00576","code":"https://github.com/OpenGalaxea/GalaxeaVLA","weights":"https://huggingface.co/OpenGalaxea/G0-VLA","project":"https://opengalaxea.github.io/GalaxeaVLA/"},"humanoid":"demonstrated","sources":[{"id":"s1","url":"https://arxiv.org/abs/2509.00576","title":"Galaxea Open-World Dataset and G0 Dual-System VLA Model","publisher":"Galaxea AI","date":"2025-08-30","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://opengalaxea.github.io/GalaxeaVLA/","title":"Original G0 project and R1 Lite hardware platform","publisher":"Galaxea AI","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://huggingface.co/OpenGalaxea/G0-VLA","title":"G0-VLA official model card","publisher":"Galaxea AI","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://raw.githubusercontent.com/OpenGalaxea/GalaxeaVLA/main/README.md","title":"GalaxeaVLA README: legacy G0 code and subsequent licensing boundaries","publisher":"Galaxea AI","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/galaxea-g0"},{"slug":"lbm-atlas","name":"Large Behavior Model for Atlas","aliases":["LBM (Atlas)"],"org":{"name":"Boston Dynamics / Toyota Research Institute","company":"boston-dynamics"},"type":"policy","released":"2025-08-20","summary":"Boston Dynamics and Toyota Research Institute train language-conditioned behavior policies that combine manipulation and locomotion on electric Atlas. The reported architecture uses a diffusion transformer with flow matching and learns from real and simulated teleoperation.","openWeights":null,"license":null,"params":450000000,"baseModel":null,"inputs":["vision","language","proprioception"],"outputs":["whole-body action chunks"],"embodiments":[{"robot":"boston-dynamics-atlas","name":"Electric Atlas","note":"Full-body policy demonstrations."},{"robot":null,"name":"Atlas Manipulation Test Stand","note":"Upper-body research platform sharing Atlas hardware and software."}],"trainingData":[{"dataset":null,"name":"Atlas and Atlas Manipulation Test Stand teleoperation"},{"dataset":null,"name":"TRI Ramen"},{"dataset":null,"name":"Simulated teleoperation"}],"links":{"paper":null,"code":null,"weights":null,"project":"https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/"},"humanoid":"native","sources":[{"id":"s1","url":"https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/","title":"Large Behavior Models and Atlas Find New Footing","publisher":"Boston Dynamics / Toyota Research Institute","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://newsroom.toyota.eu/ai-powered-robot-by-boston-dynamics-and-toyota-research-institute-takes-a-key-step-towards-general-purpose-humanoids/","title":"Boston Dynamics and Toyota advance general-purpose AI robot","publisher":"Toyota","date":"2025-08-22","type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/lbm-atlas"},{"slug":"genie-envisioner","name":"Genie Envisioner 1.0","aliases":["GE","GE-Base","GE-Act","GE-Sim"],"org":{"name":"AgiBot Genie Team / LV-NUS Lab / Beihang University","company":"agibot"},"type":"world-model","released":"2025-08-07","summary":"Genie Envisioner combines a video world model with GE-Act action prediction, demonstrated on AgiBot G1, dual Franka and Agilex platforms. The license field covers GE-Base's inherited LTX-Video 0.9 RAIL-M License; repository code uses CC BY-NC-SA 4.0 with Apache-2.0 exceptions.","openWeights":true,"license":"RAIL-M License","params":null,"baseModel":"LTX-Video 2B (original GE-Base)","inputs":["vision","language"],"outputs":["predicted multi-view video","robot actions through GE-Act"],"embodiments":[{"robot":"agibot-genie-g1","name":"AgiBot G1","note":"Real manipulation is demonstrated with the GE-Act extension, not with the video predictor alone."},{"robot":null,"name":"Dual Franka","note":"Cross-embodiment adaptation of GE-Act."},{"robot":null,"name":"Agilex Cobot Magic","note":"Cross-embodiment adaptation of GE-Act."}],"trainingData":[{"dataset":"agibot-world","name":"AgiBot World Beta"}],"links":{"paper":"https://arxiv.org/abs/2508.05635","code":"https://github.com/AgibotTech/Genie-Envisioner-V1","weights":"https://huggingface.co/agibot-world/Genie-Envisioner-v1.0","project":null},"humanoid":"demonstrated","sources":[{"id":"s1","url":"https://arxiv.org/abs/2508.05635","title":"Genie Envisioner: A Unified World Foundation Platform for Robotic Manipulation","publisher":"Genie Envisioner research team","date":"2025-08-07","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://arxiv.org/html/2508.05635v2","title":"Genie Envisioner paper: GE-Base, training data and GE-Act robot experiments","publisher":"Genie Envisioner research team","date":"2025-10-17","type":"paper","accessed":"2026-10-13"},{"id":"s3","url":"https://huggingface.co/agibot-world/Genie-Envisioner-v1.0/raw/main/README.md","title":"Genie Envisioner 1.0 checkpoint card and weight-license scope","publisher":"AgiBot World team","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://huggingface.co/Lightricks/LTX-Video/blob/main/ltx-video-2b-v0.9.license.txt","title":"LTX Video 0.9 RAIL-M License","publisher":"Lightricks","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s5","url":"https://github.com/AgibotTech/Genie-Envisioner-V1","title":"Genie Envisioner official repository: code exceptions and CC BY-NC-SA terms","publisher":"AgiBot","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/genie-envisioner"},{"slug":"gemini-robotics-on-device","name":"Gemini Robotics On-Device","aliases":[],"org":{"name":"Google DeepMind","company":null},"type":"vla","released":"2025-06-24","summary":"Gemini Robotics On-Device is a VLA designed for local robot inference and task adaptation. The announcement demonstrates ALOHA, bimanual Franka and Apptronik Apollo manipulation and offers access through a trusted-tester program.","openWeights":null,"license":null,"params":null,"baseModel":null,"inputs":["vision","language"],"outputs":["robot actions"],"embodiments":[{"robot":null,"name":"ALOHA","note":"Training embodiment named in the announcement."},{"robot":null,"name":"Bimanual Franka FR3","note":"Adapted embodiment."},{"robot":"apptronik-apollo","name":"Apptronik Apollo","note":"Humanoid adaptation demonstrated in the announcement."}],"trainingData":[{"dataset":null,"name":"ALOHA robot demonstrations"}],"links":{"paper":null,"code":null,"weights":null,"project":"https://deepmind.google/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/"},"humanoid":"demonstrated","sources":[{"id":"s1","url":"https://deepmind.google/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/","title":"Gemini Robotics On-Device brings AI to local robotic devices","publisher":"Google DeepMind","date":"2025-06-24","type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/gemini-robotics-on-device"},{"slug":"v-jepa-2-ac","name":"V-JEPA 2-AC","aliases":[],"org":{"name":"Meta FAIR","company":null},"type":"world-model","released":"2025-06-11","summary":"V-JEPA 2-AC post-trains an action-conditioned latent world model using DROID data and enables image-goal planning on Franka arms. Repository code uses MIT with Apache-2.0 exceptions; independently stated terms for the action-conditioned checkpoint were not established.","openWeights":true,"license":null,"params":null,"baseModel":"V-JEPA 2 ViT-g/16","inputs":["vision","proprioception","actions"],"outputs":["predicted visual representations"],"embodiments":[{"robot":null,"name":"Franka robot arm","note":"Real-robot planning experiments documented in the repository."}],"trainingData":[{"dataset":"droid","name":"DROID"}],"links":{"paper":"https://arxiv.org/abs/2506.09985","code":"https://github.com/facebookresearch/vjepa2","weights":"https://github.com/facebookresearch/vjepa2","project":null},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2506.09985","title":"V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning","publisher":"Meta FAIR","date":"2025-06-11","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://raw.githubusercontent.com/facebookresearch/vjepa2/main/README.md","title":"V-JEPA 2 official README: action-conditioned checkpoint and Franka planning","publisher":"Meta FAIR","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://arxiv.org/html/2506.09985v1","title":"V-JEPA 2 paper: action-conditioned model inputs and robot planning","publisher":"Meta FAIR","date":"2025-06-11","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/v-jepa-2-ac"},{"slug":"smolvla","name":"SmolVLA","aliases":["SmolVLA-450M"],"org":{"name":"Hugging Face","company":"hugging-face"},"type":"vla","released":"2025-06-02","summary":"SmolVLA combines a compact vision-language backbone with a flow-matching action expert. Hugging Face trains it on community-shared LeRobot data and reports simulated and real-arm manipulation experiments on SO100 and SO101.","openWeights":true,"license":"Apache-2.0","params":450000000,"baseModel":"SmolVLM2-500M-Video-Instruct (partial backbone)","inputs":["vision","language","proprioception"],"outputs":["continuous action chunks"],"embodiments":[{"robot":null,"name":"SO100","note":null},{"robot":null,"name":"SO101","note":null}],"trainingData":[{"dataset":null,"name":"LeRobot community SO100 manipulation datasets"}],"links":{"paper":"https://arxiv.org/abs/2506.01844","code":"https://github.com/huggingface/lerobot","weights":"https://huggingface.co/lerobot/smolvla_base","project":"https://huggingface.co/blog/smolvla"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2506.01844","title":"SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics","publisher":"Hugging Face","date":"2025-06-02","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://huggingface.co/blog/smolvla","title":"SmolVLA: architecture, community data and experiments","publisher":"Hugging Face","date":"2025-06-03","type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://huggingface.co/lerobot/smolvla_base/raw/main/README.md","title":"SmolVLA base model card","publisher":"Hugging Face","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/smolvla"},{"slug":"hume","name":"Hume","aliases":[],"org":{"name":"Shanghai AI Laboratory / Hume collaboration","company":null},"type":"vla","released":"2025-05-27","summary":"Hume combines value-guided VLA action selection with a faster reactive denoising policy, demonstrated on WidowX, Franka Panda and AgiBot G-1. The repository uses MIT for its own material and retains separate terms for inherited code and models; a complete checkpoint license is not established.","openWeights":true,"license":null,"params":null,"baseModel":null,"inputs":["vision","language","proprioception"],"outputs":["continuous action chunks"],"embodiments":[{"robot":null,"name":"WidowX 250s","note":null},{"robot":null,"name":"Franka Emika Panda","note":null},{"robot":null,"name":"AgiBot G-1","note":"Project spelling; the exact catalog generation is not unambiguously identified."}],"trainingData":[{"dataset":"libero-demos","name":"LIBERO"},{"dataset":null,"name":"BridgeData (bridge_orig LeRobot conversion; version not established)"},{"dataset":null,"name":"Fractal"}],"links":{"paper":"https://arxiv.org/abs/2505.21432","code":"https://github.com/hume-vla/hume","weights":"https://github.com/hume-vla/hume","project":"https://hume-vla.github.io/"},"humanoid":"demonstrated","sources":[{"id":"s1","url":"https://arxiv.org/abs/2505.21432","title":"Hume: Introducing System-2 Thinking in Visual-Language-Action Model","publisher":"Hume research team","date":"2025-05-27","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://hume-vla.github.io/","title":"Hume project: method and robot experiments","publisher":"Hume research team","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://github.com/hume-vla/hume","title":"Hume official repository and checkpoint instructions","publisher":"Hume research team","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://arxiv.org/html/2505.21432v4","title":"Hume paper: affiliations, model architecture, training and robot experiments","publisher":"Hume collaboration","date":"2025-07-08","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/hume"},{"slug":"gr00t-n1-5","name":"Isaac GR00T N1.5","aliases":["GR00T N1.5"],"org":{"name":"NVIDIA","company":null},"type":"vla","released":"2025-05-18","summary":"GR00T N1.5 uses a grounding-tuned Eagle 2.5 backbone and a flow-matching action transformer, with real-robot evaluations on Fourier GR-1 and Unitree G1. Its published checkpoint is labeled 3B; weights use NVIDIA OneWay Noncommercial License terms and code uses Apache-2.0.","openWeights":true,"license":"NVIDIA OneWay Noncommercial License","params":3000000000,"baseModel":"NVIDIA Eagle 2.5 (grounding-tuned variant)","inputs":["vision","language","proprioception"],"outputs":["continuous action chunks"],"embodiments":[{"robot":"fourier-gr-1","name":"Fourier GR-1","note":"Real-robot language-following evaluations."},{"robot":"unitree-g1","name":"Unitree G1","note":"Post-trained on G1 teleoperation data and evaluated on seen and unseen objects."}],"trainingData":[{"dataset":"open-x-embodiment","name":"Open X-Embodiment"},{"dataset":"agibot-world","name":"AgiBot World Beta"},{"dataset":null,"name":"Internal Fourier GR-1 demonstrations, simulated GR-1 data and DreamGen neural trajectories"}],"links":{"paper":null,"code":"https://github.com/NVIDIA/Isaac-GR00T","weights":"https://huggingface.co/nvidia/GR00T-N1.5-3B","project":"https://research.nvidia.com/labs/gear/gr00t-n1_5/"},"humanoid":"native","sources":[{"id":"s1","url":"https://huggingface.co/nvidia/GR00T-N1.5-3B","title":"GR00T-N1.5-3B model card","publisher":"NVIDIA","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://nvidianews.nvidia.com/news/nvidia-powers-humanoid-robot-industry-with-cloud-to-robot-computing-platforms-for-physical-ai","title":"NVIDIA Powers Humanoid Robot Industry With Cloud-to-Robot Computing Platforms for Physical AI","publisher":"NVIDIA","date":"2025-05-18","type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://research.nvidia.com/labs/gear/gr00t-n1_5/","title":"GR00T N1.5: model, training mixture and real-robot evaluations","publisher":"NVIDIA","date":"2025-06-11","type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://github.com/NVIDIA/Isaac-GR00T","title":"Isaac GR00T official repository: model versions and code licensing","publisher":"NVIDIA","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s5","url":"https://developer.download.nvidia.com/licenses/NVIDIA-OneWay-Noncommercial-License-22Mar2022.pdf","title":"NVIDIA OneWay Noncommercial License (22 March 2022)","publisher":"NVIDIA","date":"2022-03-22","type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/gr00t-n1-5"},{"slug":"pi0-5","name":"π0.5","aliases":["pi05","pi0.5","π₀.₅"],"org":{"name":"Physical Intelligence","company":null},"type":"vla","released":"2025-04-22","summary":"π0.5 extends π0 with heterogeneous co-training for household manipulation in unseen homes. Repository code uses Apache-2.0; Gemma Terms of Use apply to the inherited Gemma model components.","openWeights":true,"license":"Apache-2.0; Gemma Terms of Use","params":null,"baseModel":"π0","inputs":["vision","language"],"outputs":["continuous action chunks"],"embodiments":[{"robot":null,"name":"DROID Franka setup","note":"The public openpi distribution includes a DROID fine-tuned checkpoint."}],"trainingData":[{"dataset":null,"name":"Physical Intelligence multi-robot demonstrations and heterogeneous co-training data"}],"links":{"paper":"https://arxiv.org/abs/2504.16054","code":"https://github.com/Physical-Intelligence/openpi","weights":"https://github.com/Physical-Intelligence/openpi","project":null},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2504.16054","title":"π0.5: a Vision-Language-Action Model with Open-World Generalization","publisher":"Physical Intelligence","date":"2025-04-22","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://github.com/Physical-Intelligence/openpi","title":"openpi: models, checkpoints and training examples","publisher":"Physical Intelligence","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://raw.githubusercontent.com/Physical-Intelligence/openpi/main/LICENSE","title":"openpi Apache License 2.0","publisher":"Physical Intelligence","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://raw.githubusercontent.com/Physical-Intelligence/openpi/main/LICENSE_GEMMA.txt","title":"Gemma Terms of Use distributed with openpi","publisher":"Physical Intelligence / Google","date":"2024-02-21","type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/pi0-5"},{"slug":"gr00t-n1","name":"Isaac GR00T N1","aliases":["GR00T N1"],"org":{"name":"NVIDIA","company":null},"type":"vla","released":"2025-03-17","summary":"GR00T N1 is a humanoid VLA with an Eagle-2 vision-language backbone and a flow-matching action transformer; its paper reports 2.2B total parameters for GR00T-N1-2B. The weights use NVIDIA OneWay Noncommercial License terms, while repository code uses Apache-2.0.","openWeights":true,"license":"NVIDIA OneWay Noncommercial License","params":2200000000,"baseModel":"NVIDIA Eagle-2 (SmolLM2 and SigLIP-2)","inputs":["vision","language","proprioception"],"outputs":["continuous action chunks"],"embodiments":[{"robot":"fourier-gr-1","name":"Fourier GR-1","note":"Real-robot manipulation experiments in the GR00T N1 paper."},{"robot":"unitree-g1","name":"Unitree G1","note":"GR00T N1 post-trained on G1 and evaluated in NVIDIA's N1.5 research comparison."}],"trainingData":[{"dataset":"open-x-embodiment","name":"Open X-Embodiment subset"},{"dataset":null,"name":"Fourier GR-1 teleoperation and synthetic trajectories"},{"dataset":"agibot-world-alpha","name":"AgiBot-Alpha"}],"links":{"paper":"https://arxiv.org/abs/2503.14734","code":"https://github.com/NVIDIA/Isaac-GR00T","weights":"https://huggingface.co/nvidia/GR00T-N1-2B","project":null},"humanoid":"native","sources":[{"id":"s1","url":"https://huggingface.co/nvidia/GR00T-N1-2B","title":"GR00T-N1-2B model card","publisher":"NVIDIA","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://arxiv.org/html/2503.14734v1","title":"GR00T N1: An Open Foundation Model for Generalist Humanoid Robots","publisher":"NVIDIA","date":"2025-03-18","type":"paper","accessed":"2026-10-13"},{"id":"s3","url":"https://research.nvidia.com/labs/gear/gr00t-n1_5/","title":"GR00T N1.5 research: N1/N1.5 evaluations on GR-1 and G1","publisher":"NVIDIA","date":"2025-06-11","type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://github.com/NVIDIA/Isaac-GR00T","title":"Isaac GR00T official repository: model versions and code licensing","publisher":"NVIDIA","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s5","url":"https://developer.download.nvidia.com/licenses/NVIDIA-OneWay-Noncommercial-License-22Mar2022.pdf","title":"NVIDIA OneWay Noncommercial License (22 March 2022)","publisher":"NVIDIA","date":"2022-03-22","type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/gr00t-n1"},{"slug":"gemini-robotics","name":"Gemini Robotics","aliases":[],"org":{"name":"Google DeepMind","company":null},"type":"vla","released":"2025-03-12","summary":"Gemini Robotics adds robot action outputs to Gemini 2.0. It is trained primarily on ALOHA 2 data, with demonstrations on bimanual Franka arms and an adapted Apptronik Apollo humanoid.","openWeights":null,"license":null,"params":null,"baseModel":"Gemini 2.0","inputs":["vision","language"],"outputs":["robot motor commands"],"embodiments":[{"robot":null,"name":"ALOHA 2","note":null},{"robot":null,"name":"Bimanual Franka platform","note":null},{"robot":"apptronik-apollo","name":"Apptronik Apollo","note":"Humanoid adaptation demonstrated in the announcement."}],"trainingData":[{"dataset":null,"name":"ALOHA 2 robot demonstrations"}],"links":{"paper":"https://arxiv.org/abs/2503.20020","code":null,"weights":null,"project":"https://deepmind.google/blog/gemini-robotics-brings-ai-into-the-physical-world/"},"humanoid":"demonstrated","sources":[{"id":"s1","url":"https://deepmind.google/blog/gemini-robotics-brings-ai-into-the-physical-world/","title":"Gemini Robotics brings AI into the physical world","publisher":"Google DeepMind","date":"2025-03-12","type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://arxiv.org/abs/2503.20020","title":"Gemini Robotics: Bringing AI into the Physical World","publisher":"Google DeepMind","date":"2025-03-25","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/gemini-robotics"},{"slug":"gemini-robotics-er","name":"Gemini Robotics-ER","aliases":[],"org":{"name":"Google DeepMind","company":null},"type":"foundation-model","released":"2025-03-12","summary":"Gemini Robotics-ER is an embodied reasoning model for perception, spatial understanding, planning and code generation. Developers can connect its outputs to existing robot controllers.","openWeights":null,"license":null,"params":null,"baseModel":"Gemini 2.0","inputs":["vision","language"],"outputs":["spatial predictions","task plans","controller code"],"embodiments":[],"trainingData":[],"links":{"paper":"https://arxiv.org/abs/2503.20020","code":null,"weights":null,"project":"https://deepmind.google/blog/gemini-robotics-brings-ai-into-the-physical-world/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://deepmind.google/blog/gemini-robotics-brings-ai-into-the-physical-world/","title":"Gemini Robotics brings AI into the physical world","publisher":"Google DeepMind","date":"2025-03-12","type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://arxiv.org/abs/2503.20020","title":"Gemini Robotics: Bringing AI into the Physical World","publisher":"Google DeepMind","date":"2025-03-25","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/gemini-robotics-er"},{"slug":"openvla-oft","name":"OpenVLA-OFT","aliases":["OpenVLA-OFT+"],"org":{"name":"Stanford University","company":null},"type":"vla","released":"2025-02-27","summary":"OpenVLA-OFT fine-tunes OpenVLA with parallel decoding, action chunking and continuous-action regression; OFT+ adds FiLM conditioning for ALOHA. The authors describe a 7B VLA, with MIT code/checkpoint terms and inherited Llama 2 Community License Agreement restrictions from OpenVLA.","openWeights":true,"license":"MIT; Llama 2 Community License Agreement","params":7000000000,"baseModel":"OpenVLA 7B","inputs":["vision","language","proprioception"],"outputs":["continuous robot action chunks"],"embodiments":[{"robot":null,"name":"ALOHA","note":"Real-robot bimanual evaluation uses the OFT+ variant."},{"robot":null,"name":"Franka robot arm (LIBERO simulation)","note":"LIBERO-Spatial/Goal/Object/Long checkpoint variants."}],"trainingData":[{"dataset":"libero-demos","name":"LIBERO demonstrations (LIBERO checkpoint variants)"},{"dataset":null,"name":"ALOHA bimanual task demonstrations (OFT+)"},{"dataset":"open-x-embodiment","name":"Open X-Embodiment (inherited OpenVLA pretraining)"}],"links":{"paper":"https://arxiv.org/abs/2502.19645","code":"https://github.com/moojink/openvla-oft","weights":"https://huggingface.co/moojink/openvla-7b-oft-finetuned-libero-spatial","project":"https://openvla-oft.github.io/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2502.19645","title":"Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success","publisher":"Stanford University","date":"2025-02-27","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://openvla-oft.github.io/","title":"OpenVLA-OFT project: 7B VLA, OFT+ and ALOHA evaluations","publisher":"Stanford University","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://github.com/moojink/openvla-oft","title":"OpenVLA-OFT official code and pretrained-checkpoint instructions","publisher":"Stanford University","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://huggingface.co/moojink/openvla-7b-oft-finetuned-libero-spatial","title":"OpenVLA-OFT LIBERO-Spatial model card and MIT tag","publisher":"Moo Jin Kim","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s5","url":"https://github.com/openvla/openvla","title":"OpenVLA repository: inherited Llama 2 restrictions and OXE pretraining","publisher":"OpenVLA Collaboration","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/openvla-oft"},{"slug":"helix","name":"Helix","aliases":[],"org":{"name":"Figure AI","company":"figure-ai"},"type":"vla","released":"2025-02-20","summary":"Helix couples a vision-language system with a faster visuomotor policy for continuous humanoid upper-body control. Figure demonstrates household object handling and collaborative manipulation, with later logistics updates on Figure 02 and Helix-enabled Figure 03 hardware.","openWeights":null,"license":null,"params":null,"baseModel":null,"inputs":["vision","language","proprioception"],"outputs":["continuous upper-body joint commands"],"embodiments":[{"robot":"figure-02","name":"Figure 02","note":"Figure's June 2025 logistics update names this robot."},{"robot":"figure-03","name":"Figure 03","note":"The October 2025 robot announcement names Helix; Helix 02 is cataloged separately."}],"trainingData":[{"dataset":null,"name":"Figure teleoperated robot demonstrations"}],"links":{"paper":null,"code":null,"weights":null,"project":"https://www.figure.ai/news/helix"},"humanoid":"native","sources":[{"id":"s1","url":"https://www.figure.ai/news/helix","title":"Helix: A Vision-Language-Action Model for Generalist Humanoid Control","publisher":"Figure AI","date":"2025-02-20","type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://www.figure.ai/news/scaling-helix-logistics","title":"Scaling Helix: humanoid logistics","publisher":"Figure AI","date":"2025-06-07","type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://www.figure.ai/news/introducing-figure-03","title":"Introducing Figure 03","publisher":"Figure AI","date":"2025-10-09","type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/helix"},{"slug":"magma","name":"Magma","aliases":["Magma-8B"],"org":{"name":"Microsoft Research","company":null},"type":"foundation-model","released":"2025-02-18","summary":"Magma is a multimodal agent model for visual understanding and goal-directed actions in digital interfaces and robot manipulation. It uses marked objects and motion traces during training; the paper reports 8.6B total parameters for Magma-8B and real-robot WidowX rollouts.","openWeights":true,"license":"MIT","params":8600000000,"baseModel":"LLaMA 3 8B with ConvNeXt visual encoder","inputs":["vision","video","language"],"outputs":["text","visual action plans","robot actions","interface actions"],"embodiments":[{"robot":null,"name":"WidowX","note":"Real-robot manipulation rollouts on the project page."}],"trainingData":[{"dataset":"open-x-embodiment","name":"Open X-Embodiment with visual traces"},{"dataset":null,"name":"Instructional video and interface-agent datasets"}],"links":{"paper":"https://arxiv.org/abs/2502.13130","code":"https://github.com/microsoft/Magma","weights":"https://huggingface.co/microsoft/Magma-8B","project":"https://microsoft.github.io/Magma/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2502.13130","title":"Magma: A Foundation Model for Multimodal AI Agents","publisher":"Microsoft Research","date":"2025-02-18","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://huggingface.co/microsoft/Magma-8B","title":"Magma-8B official model card","publisher":"Microsoft Research","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://microsoft.github.io/Magma/","title":"Magma project: robot rollouts and training method","publisher":"Microsoft Research","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://github.com/microsoft/Magma","title":"Magma official repository and data releases","publisher":"Microsoft Research","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s5","url":"https://arxiv.org/html/2502.13130v1","title":"Magma paper: model size and backbone","publisher":"Microsoft Research","date":"2025-02-18","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/magma"},{"slug":"dexvla","name":"DexVLA","aliases":[],"org":{"name":"Midea Group / East China Normal University / Shanghai University","company":"midea-group"},"type":"vla","released":"2025-02-09","summary":"DexVLA attaches a separately pretrained diffusion action expert to a vision-language backbone. The paper studies an embodiment curriculum across single-arm, bimanual and dexterous-hand manipulation; the repository publishes code and pretrained action-expert components.","openWeights":null,"license":null,"params":null,"baseModel":"Qwen2-VL-2B","inputs":["vision","language"],"outputs":["continuous robot action chunks"],"embodiments":[],"trainingData":[],"links":{"paper":"https://arxiv.org/abs/2502.05855","code":"https://github.com/midea-ai/DexVLA","weights":null,"project":null},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2502.05855","title":"DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control","publisher":"DexVLA research team","date":"2025-02-09","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://github.com/midea-ai/DexVLA","title":"DexVLA official repository: affiliations, backbone and action-expert release","publisher":"Midea AI","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/dexvla"},{"slug":"pi0-fast","name":"π0-FAST","aliases":["pi0-FAST"],"org":{"name":"Physical Intelligence","company":null},"type":"vla","released":"2025-01-16","summary":"π0-FAST is an autoregressive VLA using frequency-domain action tokenization, with base and DROID fine-tuned checkpoints in openpi. Repository code uses Apache-2.0; Gemma Terms of Use apply to the inherited Gemma model components.","openWeights":true,"license":"Apache-2.0; Gemma Terms of Use","params":null,"baseModel":"PaliGemma-3B","inputs":["vision","language"],"outputs":["tokenized action sequences"],"embodiments":[{"robot":null,"name":"DROID Franka setup","note":"The public distribution includes pi0-FAST-DROID."}],"trainingData":[{"dataset":null,"name":"Physical Intelligence robot demonstration mixture"}],"links":{"paper":"https://arxiv.org/abs/2501.09747","code":"https://github.com/Physical-Intelligence/openpi","weights":"https://github.com/Physical-Intelligence/openpi","project":null},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2501.09747","title":"FAST: Efficient Action Tokenization for Vision-Language-Action Models","publisher":"Physical Intelligence","date":"2025-01-16","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://github.com/Physical-Intelligence/openpi","title":"openpi: models, checkpoints and training examples","publisher":"Physical Intelligence","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://raw.githubusercontent.com/Physical-Intelligence/openpi/main/LICENSE","title":"openpi Apache License 2.0","publisher":"Physical Intelligence","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://raw.githubusercontent.com/Physical-Intelligence/openpi/main/LICENSE_GEMMA.txt","title":"Gemma Terms of Use distributed with openpi","publisher":"Physical Intelligence / Google","date":"2024-02-21","type":"official","accessed":"2026-10-13"},{"id":"s5","url":"https://arxiv.org/html/2501.09747v1","title":"FAST paper: PaliGemma-based autoregressive VLA architecture","publisher":"Physical Intelligence","date":"2025-01-16","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/pi0-fast"},{"slug":"cosmos-1-0-diffusion-7b-text2world","name":"Cosmos 1.0 Diffusion 7B Text2World","aliases":["Cosmos-1.0-Diffusion-7B-Text2World"],"org":{"name":"NVIDIA","company":null},"type":"world-model","released":"2025-01-07","summary":"This Cosmos 1.0 checkpoint generates video worlds from text using a diffusion transformer. It is a world-generation model for physical AI development; the parameter count and modalities here apply specifically to the 7B Text2World variant.","openWeights":true,"license":"NVIDIA Open Model License","params":7000000000,"baseModel":null,"inputs":["language"],"outputs":["generated world video"],"embodiments":[],"trainingData":[],"links":{"paper":"https://arxiv.org/abs/2501.03575","code":null,"weights":"https://huggingface.co/nvidia/Cosmos-1.0-Diffusion-7B-Text2World","project":null},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2501.03575","title":"Cosmos World Foundation Model Platform for Physical AI","publisher":"NVIDIA","date":"2025-01-07","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://huggingface.co/nvidia/Cosmos-1.0-Diffusion-7B-Text2World","title":"Cosmos 1.0 Diffusion 7B Text2World official model card","publisher":"NVIDIA","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/cosmos-1-0-diffusion-7b-text2world"},{"slug":"cogact","name":"CogACT","aliases":[],"org":{"name":"Microsoft Research Asia / CogACT team","company":null},"type":"vla","released":"2024-11-29","summary":"CogACT separates visual-language cognition from diffusion-based action generation. It is pretrained on an Open X-Embodiment subset and evaluated on simulated Google and WidowX robots and real Realman and Franka arms.","openWeights":true,"license":"MIT","params":null,"baseModel":"Prismatic-7B (Llama 2, DINOv2 and SigLIP)","inputs":["vision","language"],"outputs":["continuous end-effector action chunks"],"embodiments":[{"robot":null,"name":"Realman robot arm","note":"Real-robot experiments; no exact catalog model is identified."},{"robot":null,"name":"Franka robot arm","note":"Real-robot experiments."},{"robot":null,"name":"Google robot","note":"SIMPLER simulation."},{"robot":null,"name":"WidowX","note":"SIMPLER simulation."}],"trainingData":[{"dataset":"open-x-embodiment","name":"Open X-Embodiment subset"}],"links":{"paper":"https://arxiv.org/abs/2411.19650","code":"https://github.com/microsoft/CogACT","weights":"https://huggingface.co/CogACT/CogACT-Base","project":"https://cogact.github.io/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://cogact.github.io/","title":"CogACT project: architecture and robot evaluations","publisher":"CogACT team","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://huggingface.co/CogACT/CogACT-Base","title":"CogACT-Base official model card","publisher":"CogACT team","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://github.com/microsoft/CogACT","title":"CogACT official repository and initial release date","publisher":"Microsoft Research","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://arxiv.org/abs/2411.19650","title":"CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulation","publisher":"CogACT team","date":"2024-11-29","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/cogact"},{"slug":"pi0","name":"π0","aliases":["pi0","pi-zero"],"org":{"name":"Physical Intelligence","company":null},"type":"vla","released":"2024-10-31","summary":"π0 combines PaliGemma with a flow-matching action expert; its paper reports 3.3B parameters. Repository code uses Apache-2.0; Gemma Terms of Use apply to the inherited Gemma model components.","openWeights":true,"license":"Apache-2.0; Gemma Terms of Use","params":3300000000,"baseModel":"PaliGemma 3B","inputs":["vision","language","proprioception"],"outputs":["continuous action chunks"],"embodiments":[{"robot":null,"name":"ALOHA","note":"Fine-tuned ALOHA checkpoints are published in openpi."},{"robot":null,"name":"DROID Franka setup","note":"Fine-tuned DROID checkpoint is published in openpi."}],"trainingData":[{"dataset":"open-x-embodiment","name":"Open X-Embodiment subset (OXE Magic Soup)"},{"dataset":null,"name":"Physical Intelligence robot demonstrations"},{"dataset":"droid","name":"DROID"},{"dataset":"bridgedata-v2","name":"BridgeData V2"}],"links":{"paper":"https://arxiv.org/abs/2410.24164","code":"https://github.com/Physical-Intelligence/openpi","weights":"https://github.com/Physical-Intelligence/openpi","project":null},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2410.24164","title":"π0: A Vision-Language-Action Flow Model for General Robot Control","publisher":"Physical Intelligence","date":"2024-10-31","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://arxiv.org/html/2410.24164v4","title":"π0 paper: architecture and training data","publisher":"Physical Intelligence","date":"2026-01-08","type":"paper","accessed":"2026-10-13"},{"id":"s3","url":"https://github.com/Physical-Intelligence/openpi","title":"openpi: models, checkpoints and training examples","publisher":"Physical Intelligence","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://raw.githubusercontent.com/Physical-Intelligence/openpi/main/LICENSE","title":"openpi Apache License 2.0","publisher":"Physical Intelligence","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s5","url":"https://raw.githubusercontent.com/Physical-Intelligence/openpi/main/LICENSE_GEMMA.txt","title":"Gemma Terms of Use distributed with openpi","publisher":"Physical Intelligence / Google","date":"2024-02-21","type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/pi0"},{"slug":"hover","name":"HOVER","aliases":["Humanoid Versatile Controller"],"org":{"name":"NVIDIA / collaborating universities","company":null},"type":"whole-body-controller","released":"2024-10-28","summary":"HOVER distills humanoid command modes into a whole-body policy with motion imitation, with Unitree H1 training and deployment code. Its code uses Apache-2.0; a separately licensed downloadable policy checkpoint was not established.","openWeights":null,"license":null,"params":null,"baseModel":null,"inputs":["proprioception","kinematic motion commands","control-mode mask"],"outputs":["whole-body motor commands"],"embodiments":[{"robot":"unitree-h1","name":"Unitree H1","note":"The repository's sim-to-real wrapper supports H1."}],"trainingData":[{"dataset":null,"name":"AMASS human motion, retargeted to Unitree H1"}],"links":{"paper":"https://arxiv.org/abs/2410.21229","code":"https://github.com/NVlabs/HOVER/","weights":null,"project":"https://hover-versatile-humanoid.github.io/"},"humanoid":"native","sources":[{"id":"s1","url":"https://arxiv.org/abs/2410.21229","title":"HOVER: Versatile Neural Whole-Body Controller for Humanoid Robots","publisher":"NVIDIA / HOVER research team","date":"2024-10-28","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://hover-versatile-humanoid.github.io/","title":"HOVER project: controller architecture and modes","publisher":"HOVER research team","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://raw.githubusercontent.com/NVlabs/HOVER/main/README.md","title":"HOVER official README: H1 deployment, AMASS data and license","publisher":"NVIDIA","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/hover"},{"slug":"rdt-1b","name":"RDT-1B","aliases":["Robotics Diffusion Transformer"],"org":{"name":"Tsinghua University (TSAIL)","company":null},"type":"vla","released":"2024-10-10","summary":"RDT-1B is a diffusion transformer for language-conditioned manipulation across heterogeneous robot action spaces. The paper reports 1.2B parameters and evaluates bimanual skills after fine-tuning on ALOHA demonstrations.","openWeights":true,"license":"MIT","params":1200000000,"baseModel":"SigLIP vision encoder and T5-v1.1-XXL language encoder","inputs":["vision","language","proprioception","control frequency"],"outputs":["continuous action chunks"],"embodiments":[{"robot":null,"name":"ALOHA","note":"Bimanual real-robot experiments."}],"trainingData":[{"dataset":"open-x-embodiment","name":"Open X-Embodiment subset"},{"dataset":"rh20t","name":"RH20T"},{"dataset":"droid","name":"DROID"},{"dataset":"bridgedata-v2","name":"BridgeData V2"},{"dataset":null,"name":"RT-1 Dataset, RoboSet and other datasets in the 46-dataset mixture"},{"dataset":null,"name":"RDT bimanual fine-tuning demonstrations"}],"links":{"paper":"https://arxiv.org/abs/2410.07864","code":"https://github.com/thu-ml/RoboticsDiffusionTransformer","weights":"https://huggingface.co/robotics-diffusion-transformer/rdt-1b","project":"https://rdt-robotics.github.io/rdt-robotics/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2410.07864","title":"RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation","publisher":"Tsinghua University","date":"2024-10-10","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://rdt-robotics.github.io/rdt-robotics/","title":"RDT-1B project: method and robot experiments","publisher":"RDT team","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://huggingface.co/robotics-diffusion-transformer/rdt-1b/raw/main/README.md","title":"RDT-1B model card: license, encoders and modalities","publisher":"RDT team","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/rdt-1b"},{"slug":"hpt","name":"Heterogeneous Pre-trained Transformers","aliases":["HPT"],"org":{"name":"MIT CSAIL / Meta FAIR","company":null},"type":"foundation-model","released":"2024-09-30","summary":"HPT pretrains a shared transformer trunk with embodiment-specific input stems and action heads; the published HPT-Base checkpoint is 12.6M parameters. Repository code uses MIT, while separately stated checkpoint license terms were not established.","openWeights":true,"license":null,"params":12600000,"baseModel":null,"inputs":["vision","proprioception"],"outputs":["shared policy representations","robot actions through task-specific heads"],"embodiments":[],"trainingData":[{"dataset":null,"name":"Heterogeneous mixture of robot teleoperation, simulation, deployed-robot data and human video"}],"links":{"paper":"https://arxiv.org/abs/2409.20537","code":"https://github.com/liruiw/HPT","weights":"https://huggingface.co/liruiw/hpt-base","project":"https://liruiw.github.io/hpt/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2409.20537","title":"Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers","publisher":"MIT CSAIL / Meta FAIR","date":"2024-09-30","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://liruiw.github.io/hpt/","title":"HPT official project page","publisher":"MIT CSAIL / Meta FAIR","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://huggingface.co/liruiw/hpt-base","title":"HPT-Base checkpoint model card","publisher":"Lirui Wang","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://github.com/liruiw/HPT","title":"HPT official code repository and MIT license","publisher":"HPT team","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/hpt"},{"slug":"tinyvla","name":"TinyVLA","aliases":[],"org":{"name":"Midea Group / East China Normal University / Shanghai University / Syracuse University","company":"midea-group"},"type":"vla","released":"2024-09-19","summary":"TinyVLA is a family of compact VLA policies that combines multimodal models with diffusion-based action generation. The authors evaluate single-arm Franka and bimanual UR5 manipulation and publish training code with several LLaVA-Pythia backbone sizes.","openWeights":null,"license":null,"params":null,"baseModel":"LLaVA-Pythia (size depends on variant)","inputs":["vision","language"],"outputs":["continuous robot actions"],"embodiments":[{"robot":null,"name":"Franka robot arm","note":"Single-arm evaluation."},{"robot":null,"name":"Bimanual UR5","note":"Bimanual evaluation."}],"trainingData":[],"links":{"paper":"https://arxiv.org/abs/2409.12514","code":"https://github.com/liyaxuanliyaxuan/TinyVLA","weights":null,"project":"https://tiny-vla.github.io/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2409.12514","title":"TinyVLA: Towards Fast, Data-Efficient Vision-Language-Action Models for Robotic Manipulation","publisher":"TinyVLA research team","date":"2024-09-19","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://tiny-vla.github.io/","title":"TinyVLA project: affiliations and robot evaluations","publisher":"TinyVLA research team","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://raw.githubusercontent.com/liyaxuanliyaxuan/TinyVLA/main/README.md","title":"TinyVLA official README: backbone variants and code release","publisher":"TinyVLA research team","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/tinyvla"},{"slug":"openvla","name":"OpenVLA","aliases":["OpenVLA-7B"],"org":{"name":"OpenVLA Collaboration","company":null},"type":"vla","released":"2024-06-13","summary":"OpenVLA adapts a Prismatic vision-language backbone to tokenized robot actions; its authors describe the policy as 7B. The code and checkpoint card specify MIT, while the repository explicitly states that pretrained models also inherit the Llama 2 Community License Agreement.","openWeights":true,"license":"MIT; Llama 2 Community License Agreement","params":7000000000,"baseModel":"Prismatic prism-dinosiglip-224px (Llama 2, DINOv2 and SigLIP)","inputs":["vision","language"],"outputs":["end-effector action tokens"],"embodiments":[{"robot":null,"name":"WidowX","note":"The model card documents BridgeData V2 zero-shot control."}],"trainingData":[{"dataset":"open-x-embodiment","name":"Open X-Embodiment mixture"}],"links":{"paper":"https://arxiv.org/abs/2406.09246","code":"https://github.com/openvla/openvla","weights":"https://huggingface.co/openvla/openvla-7b","project":"https://openvla.github.io/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2406.09246","title":"OpenVLA: An Open-Source Vision-Language-Action Model","publisher":"OpenVLA Collaboration","date":"2024-06-13","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://huggingface.co/openvla/openvla-7b","title":"OpenVLA-7B official model card","publisher":"OpenVLA Collaboration","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://github.com/openvla/openvla","title":"OpenVLA official repository: pretrained models and inherited license restrictions","publisher":"OpenVLA Collaboration","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://ai.meta.com/llama/license/","title":"Llama 2 Community License Agreement","publisher":"Meta","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/openvla"},{"slug":"octo","name":"Octo","aliases":["Octo-Base"],"org":{"name":"Octo Model Team","company":null},"type":"policy","released":"2024-05-20","summary":"Octo is a generalist transformer policy conditioned on language or goal images and trained on an Open X-Embodiment mixture. The parameter count here describes the published Octo-Base checkpoint; smaller checkpoints are also available.","openWeights":true,"license":"MIT","params":93000000,"baseModel":null,"inputs":["vision","language","goal image"],"outputs":["continuous action chunks"],"embodiments":[{"robot":null,"name":"WidowX","note":"Official repository supplies real-robot evaluation instructions."}],"trainingData":[{"dataset":"open-x-embodiment","name":"Open X-Embodiment mixture"}],"links":{"paper":"https://arxiv.org/abs/2405.12213","code":"https://github.com/octo-models/octo","weights":"https://huggingface.co/rail-berkeley/octo-base","project":"https://octo-models.github.io/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://arxiv.org/abs/2405.12213","title":"Octo: An Open-Source Generalist Robot Policy","publisher":"Octo Model Team","date":"2024-05-20","type":"paper","accessed":"2026-10-13"},{"id":"s2","url":"https://huggingface.co/rail-berkeley/octo-base","title":"Octo-Base model card","publisher":"Octo Model Team","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s3","url":"https://raw.githubusercontent.com/octo-models/octo/main/README.md","title":"Octo official README: checkpoints and evaluation","publisher":"Octo Model Team","date":null,"type":"official","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/octo"},{"slug":"rt-2","name":"RT-2","aliases":["Robotics Transformer 2"],"org":{"name":"Google DeepMind","company":null},"type":"vla","released":"2023-07-28","summary":"RT-2 co-trains pretrained vision-language models on web tasks and robot trajectories, representing robot actions as text tokens. The project evaluates PaLM-E and PaLI-X variants, so no single parameter count is assigned to the family.","openWeights":null,"license":null,"params":null,"baseModel":"PaLM-E or PaLI-X (variant-dependent)","inputs":["vision","language"],"outputs":["robot action tokens","natural-language reasoning"],"embodiments":[],"trainingData":[{"dataset":null,"name":"RT-1 robot demonstration dataset"},{"dataset":null,"name":"Web vision-language data used by PaLM-E and PaLI-X"}],"links":{"paper":"https://arxiv.org/abs/2307.15818","code":null,"weights":null,"project":"https://robotics-transformer2.github.io/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://robotics-transformer2.github.io/","title":"RT-2 project: method and model variants","publisher":"Google DeepMind","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://arxiv.org/abs/2307.15818","title":"RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control","publisher":"Google DeepMind","date":"2023-07-28","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/rt-2"},{"slug":"rt-1","name":"RT-1","aliases":["Robotics Transformer 1"],"org":{"name":"Robotics at Google","company":null},"type":"policy","released":"2022-12-13","summary":"RT-1 uses image sequences and language instructions to predict discretized arm and mobile-base controls. Its paper reports 35M parameters, and the official repository publishes pretrained checkpoints.","openWeights":true,"license":"Apache-2.0","params":35000000,"baseModel":"EfficientNet with Universal Sentence Encoder language embeddings","inputs":["vision","language"],"outputs":["discrete arm and base action tokens"],"embodiments":[],"trainingData":[{"dataset":null,"name":"RT-1 robot demonstration dataset"}],"links":{"paper":"https://arxiv.org/abs/2212.06817","code":"https://github.com/google-research/robotics_transformer","weights":"https://github.com/google-research/robotics_transformer","project":"https://robotics-transformer1.github.io/"},"humanoid":"none-reported","sources":[{"id":"s1","url":"https://robotics-transformer1.github.io/","title":"RT-1 project: architecture, data and experiments","publisher":"Robotics at Google","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s2","url":"https://arxiv.org/abs/2212.06817","title":"RT-1: Robotics Transformer for Real-World Control at Scale","publisher":"Robotics at Google","date":"2022-12-13","type":"paper","accessed":"2026-10-13"},{"id":"s3","url":"https://github.com/google-research/robotics_transformer","title":"Robotics Transformer official repository and pretrained checkpoints","publisher":"Google Research","date":null,"type":"official","accessed":"2026-10-13"},{"id":"s4","url":"https://arxiv.org/html/2212.06817v2","title":"RT-1 paper: 35M architecture and language encoder","publisher":"Robotics at Google","date":"2023-08-11","type":"paper","accessed":"2026-10-13"}],"lastVerified":"2026-10-13","url":"https://humanoidversus.com/models/rt-1"}]}