tmeynier/hepha_act_100_simple_drawer_5

LeRobot-style behavior-cloning dataset generated from the Hepha MuJoCo simulation.

Summary

• Robot type: hepha_mujoco
• Codebase version: v3.0
• Episodes: 100
• Frames: 200000
• FPS: 30
• Joint normalization: min_max_0_1

Features

• timestamp: float32 [1]
• frame_index: int64 [1]
• episode_index: int64 [1]
• index: int64 [1]
• task_index: int64 [1]
• episode.drawer_index: int64 [1]
• episode.cube_position: float32 [3]
• episode.cube_quaternion: float32 [4]
• observation.drawer_index: int64 [1]
• observation.state: float32 [15]
• action: float32 [15]
• observation.state_raw: float32 [15]
• action_raw: float32 [15]
• observation.images.head_camera: video [3, 480, 640]

Policy-Facing Columns

• observation.state: normalized robot joints in [0, 1]
• observation.drawer_index: selected drawer target for each frame
• action: normalized next-step robot joint targets in [0, 1]
• observation.images.head_camera: RGB video frames from the robot camera

Extra Columns

• observation.state_raw: raw MuJoCo joint positions
• action_raw: raw next-step MuJoCo joint positions
• episode.drawer_index: selected drawer index for each frame
• episode.cube_position: initial cube position for each frame
• episode.cube_quaternion: initial cube orientation for each frame

Notes

The dataset was produced by first computing the full IK episode, then resampling the robot joint trajectory at a constant normalized joint speed before saving frames/actions. Training code appends the normalized drawer target (drawer_index - 1) / 8 to the policy state when observation.drawer_index is present.

Joint limits and the normalization formula are stored in meta/info.json.