NoahWeiss/factory-episodes

Factory Episodes — laptop-encoded

161 bimanual manipulation episodes recorded on a nero-dual robot, laptop-encoded to H.264.

Layout

One directory per episode:

episode_<YYYYMMDD>_<HHMMSS>_<id>/
  meta.json              # episode metadata (see below)
  frames.jsonl           # one JSON object per timestep
  cameras/base.mp4       # 1280x720 RGB, 30 fps
  cameras/wrist_left.mp4
  cameras/wrist_right.mp4
  ir/base.mp4            # IR stream (when recorded)

frames.jsonl

One line per frame:

{
  "frame": 0,
  "t": 0.0,
  "mono": 5407.795774,
  "live": false,
  "left":  {"joints": [7 floats], "gripper_pct": 99.8},
  "right": {"joints": [7 floats], "gripper_pct": 99.3}
}

• t — seconds from episode start; mono — monotonic clock stamp.
• joints — 7-DoF arm joint positions in radians.
• gripper_pct — gripper opening, 0–100.

meta.json

Format nero-episode-v3. Includes fps, duration_s, frame_count, camera serials, rgb_size, per-camera video records (frames, bytes, wall_start), task, and action_source (next_observation — actions are the next observed state).

Video timestamps use the wall clock (video_clock: "wall"); align frames to video via each camera's wall_start in meta.json.

Loading

import json
from huggingface_hub import snapshot_download

root = snapshot_download("NoahWeiss/factory-episodes", repo_type="dataset")
meta = json.load(open(f"{root}/episode_20260823_045828_d5cb8c8e/meta.json"))
frames = [json.loads(l) for l in open(f"{root}/episode_20260823_045828_d5cb8c8e/frames.jsonl")]