MichaelGan0625/piston_long_horizon_videos

Piston Long-Horizon Videos

This dataset contains 107 long-horizon piston replay recordings generated from the local archive piston_long_horizon_ego_wrist_videos_20260726.tar.

Each trajectory has two synchronized camera-view videos:

• ego/: ego/front camera recordings
• wrist/: right wrist camera recordings

The videos were rendered from the 50 Hz replay pipeline used for the Piston task. The replay uses a free-physics object mode rather than strictly forcing the object pose at every frame, so the videos are intended for visual observation/replay inspection rather than as a ground-truth physical-state dataset.

Directory Structure

ego/
  trajectory_csv_1/
    episode_trial0_env0.mp4
    ...
  trajectory_csv_2/
    ...
  trajectory_csv_3/
    ...
wrist/
  trajectory_csv_1/
    episode_trial0_env0.mp4
    ...
  trajectory_csv_2/
    ...
  trajectory_csv_3/
    ...
manifest.csv

Manifest

manifest.csv contains one row per trajectory and aligns the two camera views:

Column	Description
trajectory_group	Source trajectory group, for example trajectory_csv_1.
episode	Episode identifier, for example episode_trial0_env0.
ego_video	Relative path to the ego/front camera video.
wrist_video	Relative path to the right wrist camera video.

Counts

• Trajectories: 107
• Ego/front camera videos: 107
• Right wrist camera videos: 107
• Total MP4 files: 214

Loading Example

import pandas as pd

manifest = pd.read_csv("manifest.csv")
row = manifest.iloc[0]

ego_path = row["ego_video"]
wrist_path = row["wrist_video"]

Limitations

• This dataset contains rendered replay videos only.
• It does not include the original CSV trajectories, actions, robot states, or object states.
• The recordings are simulated IsaacLab/Unitree G1 replay outputs, not real robot camera footage.
• No train/validation/test split is provided.