physicalai-bmi/forge-arm-reach

An open robot-trajectory dataset generated fully in the browser, on-device with Institute for Physical AI · Forge (in-browser MuJoCo, on-device).

• Robot: forge_arm
• Task: Reach the target with the end effector
• Episodes: 11 · Frames: 3777 · Control rate: 60 Hz
• observation.state: 7 dims · action: 3 dims

Format

LeRobot v2.1-compatible. meta/ holds info.json, episodes.jsonl, tasks.jsonl, and stats.json.

Per-episode frames are provided both ways under data/chunk-000/:

• episode_*.parquet — canonical LeRobot parquet (what info.json's data_path points at). Ready to load.
• episode_*.jsonl — the exact source produced client-side by the in-browser exporter, kept for provenance.

to_parquet.py is the (already-applied) converter, included so anyone can reproduce the parquet from the JSONL.

Load

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("physicalai-bmi/forge-arm-reach")
print(ds[0]["observation.state"], ds[0]["action"])

How it was made

Recorded in a real browser against the live Forge: the "Reach Trials" arm preset, auto-expert demonstrations, exported on-device. No hardware, no cloud, no install. Reproduce it yourself in about a minute.

See DATASHEET.md for provenance, composition, and intended uses. Licensed CC-BY-4.0.