felixfabricius/robot-chess-commentator-squares

Robot chess commentator: square dataset

Labelled square crops of a real chess board photographed by a Reachy Mini robot, labelled with the piece standing on each square. 23,744 labelled squares from 371 board positions. Was used to train a square classification model that enables the robot to commentate chess games

Configs

squares -- one row per square crop

Split	Rows
train	13,184
val	4,928
test	5,632

Column	Notes
image	144x144 RGB crop.
mask	144x144 crop mask, {0, 255}.
annotated	The crop with its hull outline drawn on.
label	13 classes: empty, KQRBNP (white), kqrbnp (black).
top, left	Crop origin within the padded warped board.
top_left_corner	Which board corner is top-left in the camera image.
board_fen, previous_board_fen, move_uci	Position context; empty for free-placement edits.
valid_game_position	False where the position was set up by hand rather than reached by legal play.
fen_matches_pixels	False on the 28 boards where the recorded FEN disagrees with the photograph. Filter on it for move evaluation; ignore it for classification.

Reconstructing the model input. The model takes 4 channels, RGB plus the mask:

import numpy as np
from datasets import load_dataset

row = load_dataset("felixfabricius/robot-chess-commentator-squares", "squares", split="test")[0]
rgb = np.array(row["image"])                    # (144, 144, 3) uint8
mask = np.array(row["mask"]) // 255             # (144, 144)    uint8, {0, 1}
model_input = np.dstack([rgb, mask])            # (144, 144, 4) uint8

Scale to [0, 1] by dividing the RGB channels by 255. There is no mean/std normalisation.

boards -- one row per captured position

warped_image is the top-down rectified board; labels is the 64-way label vector in file-major order (a1, a2, ..., a8, b1, ...). Join to squares on (setup_id, image_id).

Splits are by setup, not by square

A setup is one physical calibration determined by the position of the robot relative to the chess board. Each setup contributes multiple positions and splits are stratified by setup.

Labels

Labels were extracted from a virtual chess board that mirrored the physical chess board - see blog for details. To allow for more interesting square pictures, free placement of pieces on the virtual board was allowed. valid_game_position=False marks positions reached by such free placement.

label describes the pixels; board_fen describes the virtual game

Occasionally the virtual board drifted out of sync with the physical one. When that was noticed, the labels were corrected to match the photograph; the FEN was not. On 28 of 371 boards (52 squares) the two therefore disagree, and fen_matches_pixels is False.

Which one you want depends on the task:

• Square classification -- use label and ignore the flag.
• Move estimation -- you need previous_board_fen -> legal moves -> move_uci, and that chain describes the virtual game. Filter to fen_matches_pixels first:

boards = load_dataset("felixfabricius/robot-chess-commentator-squares", "boards", split="test")
evaluable = boards.filter(
    lambda r: r["fen_matches_pixels"] and r["valid_game_position"] and r["move_uci"]
)

Of the 217 boards that carry a recorded move at all, only one (in val) is affected.

License

CC-BY-4.0. Copyright 2026 Felix Fabricius. The code that produced it is GPL-3.0-or-later; the trained model is at felixfabricius/robot-chess-commentator-cnn.