Robot AI model
CogACT
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.
Records last verified Oct 13, 2026 · Send a correction
- Organization
- Microsoft Research Asia / CogACT team
- Type
- VLA
- Released / introduced
- Nov 29, 2024
- Public weights
- Public weights
- Licence / scope
- MIT
- Parameters
- Not disclosed
- Base model
- Prismatic-7B (Llama 2, DINOv2 and SigLIP)
- Humanoid association
- None reported
- Inputs
- vision, language
- Outputs
- continuous end-effector action chunks
Release dates can mark an announcement or paper introduction. Public checkpoints, code licences and weight terms are distinct; follow the publisher’s links for the stated scope.
Documented embodiments
Source-associated hardware; this list does not establish plug-and-play compatibility or commercial robot capability.
- Realman robot arm
Real-robot experiments; no exact catalog model is identified.
- Franka robot arm
Real-robot experiments.
- Google robot
SIMPLER simulation.
- WidowX
SIMPLER simulation.
Disclosed training data
More robot AI models
Evidence
Sources
- [s1] CogACT project: architecture and robot evaluations ↗
CogACT team · official · accessed Oct 13, 2026
- [s2] CogACT-Base official model card ↗
CogACT team · official · accessed Oct 13, 2026
- [s3] CogACT official repository and initial release date ↗
Microsoft Research · official · accessed Oct 13, 2026
- [s4] CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulation ↗
CogACT team · paper · published Nov 29, 2024 · accessed Oct 13, 2026