338 episodes · 50 fps · 2 cameras · 320×240 av1
Humanoid Chess FB32 — all five pieces, left arm
338 teleoperated demonstrations of a simulated Unitree G1 humanoid moving chess pieces with its left three-finger hand on a full 8x8 board in NVIDIA Isaac Sim 5.1. Part of the FB32 benchmark (masters research, University of the Witwatersrand).
- Pieces: king, queen, rook, bishop, knight (both colours), ~68 demos each; 34 distinct source-to-destination moves per piece, each collected twice. Only successful demos are included.
- Collection: VR teleoperation (Meta Quest 3 over NVIDIA CloudXR), 50 Hz, Pink IK on the wrists, direct hand-joint targets.
- Size: 338 episodes, LeRobot v3.0 format.
Features
| key | shape | description |
|---|---|---|
| observation.images.head | 240x320x3 video | head camera |
| observation.images.left_wrist | 240x320x3 video | left wrist camera |
| observation.state | 28 | both wrist positions (3+3) and quaternions (4+4), 14 hand joint angles |
| action | 28 | absolute wrist pose targets (both arms) + 14 hand joint targets |
| task | text | e.g. "Move the white knight from h1 to f2." |
The destination square shows a flat coloured marker whose colour encodes the piece type (king yellow, queen violet, rook red, bishop magenta, knight cyan).
Loading
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("Raysolo/fb32-v03-all5-left")
Use in the FB32 benchmark
ACT, Diffusion Policy, SmolVLA and GR00T N1.6/N1.7 were trained on this data and evaluated on four suites (trained moves, identifying the right piece among two, an unfamiliar pawn as distractor, an unfamiliar pawn as target).
- Downloads last month
- 127