Humanoid Everyday: A Comprehensive Robotic Dataset for Open-World Humanoid Manipulation
Humanoid Everyday is a real-world humanoid manipulation dataset with 10.3k trajectories and over 3 million frames across 260 tasks in 7 broad categories. It was collected on Unitree G1 and H1 humanoids and records RGB, depth, LiDAR and tactile streams with language annotations. The authors also benchmark policy-learning methods and describe a cloud evaluation platform for external policies, which the project page lists as coming soon.
Open-world humanoid manipulation across dexterous object manipulation, human-humanoid interaction and loco-manipulation. The dataset is a training resource; the planned cloud evaluation would score deployed policies on real humanoids.
The seven categories are Basic Manipulation, Deformable Manipulation, Articulated Manipulation, Tool Use, High-Precision Manipulation, Human-Robot Interaction and Loco-Manipulation. Data are recorded at 30 Hz. The paper describes a cloud platform that streams observations to a locally hosted humanoid, runs the policy on the user's machine and returns actions to the robot.
This dataset spans 260 tasks, including categories with locomotion and human-robot interaction, which are directly relevant to humanoid generalist policies. It records on two humanoid platforms with dexterous hands, so it can support comparisons across bodies.
Scoring
This record defines no leaderboard metrics.
Hardware
Collected on two Unitree humanoids: the G1 (29 DoF, with 7-DoF three-finger Dex3-1 hands with tactile sensing) and the H1 (27 DoF, with 6-DoF Inspire hands). Both carry Intel RealSense RGB-D cameras and a Livox LiDAR, according to the paper.
Recorded outcomes
No result files yet
Evidence
Related evaluations