A benchmark dataset for bimanual robot manipulation, comprising 50 distinct tasks. It was converted to the LeRobot v3.0 format by GT-111 and includes lossless stream-copied H.264 video. The dataset provides absolute pose state and action data for both robot arms.
Use Cases
- Training imitation learning models based on recorded end-effector poses and gripper actions
- Benchmarking robot policy performance across 50 distinct manipulation tasks
- Developing visual servoing algorithms based on multi-camera video streams
- Analyzing bimanual coordination strategies using absolute quaternion and position data
Strengths
- Contains 50 distinct bimanual manipulation tasks, providing a broad test suite
- Data is converted to a standardized format (LeRobot v3.0) with lossless video
- Provides 16-dimensional end-effector data per frame for both arms
Limitations
- Description metadata is limited; actual data quality requires manual inspection after download
- Column-level documentation is absent; field semantics must be inferred after download
- Row count is unknown, which may limit suitability assessment
Provenance
- Source
- GT-111 on Hugging Face
- Collection Method
- Converted from the RobotWin benchmark to LeRobot v3.0 format
- Freshness
- Last updated 2026-06-01 06:33:47; freshness should be verified