so101_chess_wide

Chess moves and captures by an SO-101 arm, recorded so that the setup does not matter: boards of different sizes and materials, twenty-two piece sets, the workspace camera anywhere from a mast above the board to a tripod beside the table, the arm clamped where it fits, and the light and the table changing from episode to episode. Scripted-expert demonstrations in MuJoCo, every one verified. The fixed-rig set this grew out of is so101_chess.

This is the second collection. In the first, kept as revision v1, the camera's place and the board's turn and offset were drawn once per 25-episode scene - some three hundred placements in seven thousand episodes - and a policy trained on it learned none of them. Here both are drawn every episode: 3,714 camera poses and 7,088 board placements in 7,732 episodes.

episodes	7,732: 5,213 moves (pick up the piece on e2 and place it on e4) and 2,519 captures (take the piece on d5 off the board)
frames	813,592 at 10 Hz
cameras	observation.images.top - the workspace camera, wherever it stands - and observation.images.wrist, 640 x 480
state and action	six joints, SO-101 degrees; the action is the expert's commanded target
format	LeRobot v3.0, plus scenes.jsonl

What varies

Drawn once per 25 episodes, because it is compiled into the scene:

	range	share
square size	25-31 mm	a fifth on the 28 mm board of the fixed-rig set
border, thickness	6-25 mm, 3-20 mm
board material	wood, plastic, vinyl, magnetic - each with its own look and grip on the pieces	37 / 23 / 21 / 20 %
coordinates printed on the border		46%
distance from the arm, arm mounting	5-11 cm; clamped to the table, or on a 2-6 cm riser	16% on the table
piece set	22 sets, below	219-550 episodes each
table	plain, wood, cloth, marble, laminate	about a fifth each

Drawn every episode:

	range	share
board placement	turned up to 12 degrees either way, up to 4 cm off the arm's centre line, plus a shift of up to 10 mm
workspace camera	on the mast of the fixed rig / on the line from the mast to above the board / on a boom arm up to 25 cm off the board centre, 35-75 cm up / beside the table, 35-60 cm out and 20-50 degrees up	33 / 21 / 31 / 16 %
light	the key light's direction, brightness and colour from a warm lamp to cool daylight, the room's fill and ambient light, how soft the shadows are
colours, start	piece and table colours, the room behind the table, the arm's start pose within 0.1 rad

Tasks come only from squares the arm can reach in that placement, so the far corners of a 31 mm board appear less often than the middle.

Piece sets

• Nineteen generated sets - maple, boxwood, birch, walnut, bamboo, bone, ivory resin, alabaster, brass, steel, frosted glass, three plastics, painted, carved, minimalist, a Soviet club style and a travel set - made text-to-3D with EmbodiedGen. Pawns, knights, bishops, queens and kings are generated; in seventeen sets the rook is this project's own rook shape in the set's colour, because the image model would not draw one. A few generated bishops are plain cones. The black side is the same mesh under a dark tint.
• Two token sets - flat discs and short cylinders with the piece's symbol on top, 24 x 9 mm and 18 x 16 mm, as travel sets have them. Tokens thinner than about 9 mm are not here: the SO-101's jaws cannot close on them.
• The set of the fixed-rig data.

Every set had to pass the scripted expert first: 18 moves and captures, at least 90% done.

Held out

Three things were kept out so that a policy trained on this can be scored on what it never saw: the marble piece set, the large token set (25 x 11 mm), and every cardboard board - about a quarter of the recorded episodes. The project's evaluator draws scenes from exactly those:

python examples/evaluate_policy.py --checkpoint <policy> --moves 32 --captures 16 --scenes 8 \
    --piece-sets marble token_disc_big --board-finishes cardboard

scenes.jsonl

One line per episode, in episode order: the instruction and move, the piece set, the board (square, border, thickness, turn, distance to the arm, riser), its material and labels, the table surface, the camera (placement, height, offset from the board centre, field of view, roll) and the light. Filter on it to train on a camera placement or a board size alone:

import json
from lerobot.datasets.lerobot_dataset import LeRobotDataset

scenes = [json.loads(line) for line in open("scenes.jsonl")]
overhead = [s["episode_index"] for s in scenes if s["camera"]["placement"] != "side"]
dataset = LeRobotDataset("XvKuoMing/so101_chess_wide", episodes=overhead)

How it was recorded

A scripted pick-and-place expert drives a calibrated SO-101 model; an episode is kept only if the piece ended where the instruction said, upright, with nothing else displaced. The expert succeeded on about 95% of attempts across the variety above, and its failed attempts were dropped. Everything is generated by the project repository: examples/collect_demonstrations.py --randomize --vary-board --piece-sets all.

Limits

• Simulation only.
• A policy trained on the fixed rig alone scores 2 of 48 on scenes drawn from this distribution. The one trained on the first collection (revision v1) scored 1 of 48 after 20,000 steps; a policy trained on this collection is in training and has no result yet.
• A side camera hides back-rank pieces behind front ones. Expect it to be the hardest placement.
• A board whose top is more than about 12 mm above the arm's base puts the board edge in the forearm's path; the recordings raise the arm for thick boards, and a real setup has to as well.