h2r-sim / README.md
batgre's picture
Add dataset card
e12b5d9 verified
|
Raw History Blame Contribute Delete
1.36 kB
metadata
license: mit
task_categories:
  - robotics
tags:
  - world-model
  - robosuite

h2r-sim

Robosuite episodes with real actions, used to train the action controller of human-to-robot (Panda action → latent action of the world model).

  • Robot: Panda, OSC_POSE controller at 20 Hz. Each action is held for 5 control steps, so one action per frame (4 Hz).
  • Camera: agentview, 224 × 224 JPEG.
  • Tasks: Lift and PickPlaceCan. Policies: scripted (noisy waypoints: reach, grasp and lift the object, or random targets in the workspace) and random (uniform random translation and gripper; pushes that leave the workspace are cancelled). Rotation is always zero.
  • 500 episodes per task and policy, 40 actions and 41 frames per episode. Episode k uses seed k, so both policies start from the same object placements.

Layout

  • frames/<split>-<shard>.parquet: one row per episode.
    • video: <task>/<policy>/episode_<seed>; task, policy
    • frames: 41 JPEG images
    • actions: 40 × 7 (dx, dy, dz, rotation × 3, gripper), in [-1, 1]
    • eef, object: 41 × 3 positions of the end effector and the object (m)
    • states: 41 flattened MuJoCo states, to reset the simulator to any frame
  • manifest.parquet: video id, split, number of frames, shard.
  • config.yaml: the collection config.