Robomimic PH: Human-Demonstrated Robot Manipulation Trajectories for Four Tasks

800 episodes of robot manipulation data collected from a proficient human operator performing four distinct tasks: lifting a cube, picking and placing a can, squaring, and tool hanging. The dataset was created using LeRobot and is formatted for training Vision-Language-Action (VLA) models like OpenPI. It was uploaded to Hugging Face by yananchen and last updated on December 5, 2025.

Use Cases

• Training imitation learning models based on proficient human demonstration trajectories.
• Benchmarking robot policy performance on the lift, can, square, and tool hang tasks.
• Developing multi-task robotic control systems based on the four included manipulation scenarios.
• Fine-tuning vision-language models for robotic instruction following based on the dataset's multimodal nature.

Strengths

• Contains 200 episodes per task for a total of 800 demonstration trajectories.
• Data is sourced exclusively from a Proficient-Human (PH) operator, which may indicate high-quality demonstrations.
• Specifically formatted for compatibility with modern VLA training pipelines like LeRobot v3.0.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown beyond the episode total, which may limit suitability assessment for large-scale training.
• The description metadata is limited; actual data structure and quality require manual inspection after download.

Provenance

Source

Created using LeRobot, based on the Robomimic framework and Robosuite simulations.

Collection Method

Trajectories collected from a Proficient-Human (PH) operator performing tasks in a simulated environment.

Time Range

null

Freshness

Last updated 2025-12-05 21:02:55; freshness should be verified.

Geography

null