cagataydev/so101-tic-tac-toe

SO-101 Tic-Tac-Toe — 210 teleoperated episodes, 20 language-labelled tasks

Two 3D-printed piece types (pink X, white O), one printed board, one SO-101 follower arm driven by an SO-101 leader. Every episode is a single atomic move — pick a piece, or pick a piece and place it on a named cell — recorded at 30 Hz with an overhead camera and a wrist camera, in the LeRobot v3.0 format.

episodes	210
frames	57,092 at 30 Hz = 31.7 min of teleoperation
tasks	20 natural-language strings (table below)
cameras	observation.images.top 480×640 · observation.images.wrist 480×640 (AV1 video)
state / action	6 floats: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll in degrees, gripper in percent
robot	SO-101 follower (so_follower), SO-101 leader for teleop, Feetech STS3215
recorded	2026-10-25/26, one session, one scene, one lighting setup
format	LeRobot v3.0 (meta/info.json, meta/episodes/*.parquet, meta/tasks.parquet)

Tasks

The strings are the ones a policy is conditioned on — use them verbatim.

task	episodes	mean s	min–max s
pick up the x piece	20	11.2	6.8–29.7
pick up the o piece	10	9.4	7.0–14.1
pick up the x piece and place it on the top left cell	10	11.4	7.0–19.4
pick up the x piece and place it on the top center cell	10	8.0	6.5–9.7
pick up the x piece and place it on the top right cell	10	7.7	6.5–10.2
pick up the x piece and place it on the middle left cell	10	8.2	6.7–9.6
pick up the x piece and place it on the center cell	10	8.4	6.4–12.1
pick up the x piece and place it on the middle right cell	10	8.0	5.9–10.2
pick up the x piece and place it on the bottom left cell	10	8.5	6.9–10.2
pick up the x piece and place it on the bottom center cell	10	8.1	6.6–9.9
pick up the x piece and place it on the bottom right cell	10	8.2	6.8–10.8
pick up the o piece and place it on the top left cell	10	10.2	7.2–13.2
pick up the o piece and place it on the top center cell	10	8.3	6.8–11.8
pick up the o piece and place it on the top right cell	10	9.5	8.1–11.4
pick up the o piece and place it on the middle left cell	10	8.0	6.3–10.4
pick up the o piece and place it on the center cell	10	10.2	7.0–13.5
pick up the o piece and place it on the middle right cell	10	8.6	7.6–10.4
pick up the o piece and place it on the bottom left cell	10	9.1	7.2–13.4
pick up the o piece and place it on the bottom center cell	10	9.0	6.5–12.1
pick up the o piece and place it on the bottom right cell	10	9.3	6.3–15.5
total	210	9.1	5.9–29.7

⚠ Cell names are in the robot's frame, not the camera's

The arm is mounted at the top edge of the overhead image, facing down the board. Cells were named from the robot's point of view — as if standing behind the arm looking at the board — so they appear rotated 180° in the top camera: the robot's top left is the camera's bottom-right, bottom center is the camera's top-center, center is center. This is consistent across all 180 placing episodes (verified frame by frame). If you name cells from the image, 8 of 9 labels will look wrong — they are not.

camera image (top camera)              robot's frame (task strings)
┌───────────────┐                      ┌───────────────┐
│  arm base     │                      │  BL   BC   BR │   BL = "bottom left" …
│ ┌───┬───┬───┐ │                      │ ┌───┬───┬───┐ │
│ │BR │BC │BL │ │   ⟲ 180°   ⇄         │ │   │   │   │ │
│ ├───┼───┼───┤ │                      │ ├───┼───┼───┤ │
│ │MR │ C │ML │ │                      │ │   │   │   │ │
│ ├───┼───┼───┤ │                      │ ├───┼───┼───┤ │
│ │TR │TC │TL │ │                      │ │   │   │   │ │
│ └───┴───┴───┘ │                      │ └───┴───┴───┘ │
│  far side     │                      │  TL   TC   TR │
└───────────────┘                      └───────────────┘

Scene

• Board: white card with a wooden 3×3 grid, centred under the overhead camera; a Strands Agents card sits in front of it as a fixed visual anchor. The board drifts a few millimetres between episodes (it is not fixed down).
• Pieces: X pile on the left of the board, O pile on the right (camera frame). The operator varies which piece is grabbed and its orientation. At the start of every episode the board is empty.
• Cameras: top — Logitech 4K Pro looking straight down; wrist — USB camera on the gripper, looking along the fingers. At the end of most episodes the arm retracts and the wrist camera sees the operator's monitor (live terminal text) — a moving distractor a policy should learn to ignore.
• Lighting: a single desk lamp from the right of the board, switched on from episode 10 onward (episodes 0–9 of pick up the x piece are darker, especially in the wrist view).
• Episode boundaries: the operator ends each episode with the right-arrow key the moment the piece is released, so there is little dead time; episodes are 5.9–29.7 s (median ≈ 8.5 s). After each episode the piece goes back to its pile.

Recording

Teleoperation with lerobot-record (lerobot main, Oct 2025) through a thin wrapper that records a plan file of tasks back to back, 10 episodes each, resumable — both files ship with the dataset: recording/record.py and recording/tasks.txt. Cadence over the whole session: 29.97 Hz effective, 0 of 2,775 ticks over the 33.3 ms budget in the final run, no dropped frames (see meta/episodes lengths vs. timestamps). Videos are SVT-AV1 CRF 30, yuv420p, GOP 2, encoded after each episode.

Pieces and board: Hashtag-Robotics/so101-tic-tac-toe (printable X/O pieces and the grid). Our task strings differ from the HashtagRobotics datasets on purpose — ours name a single move.

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("cagataydev/so101-tic-tac-toe")
ds.meta.tasks                      # the 20 strings ↔ task_index
sample = ds[0]
sample["observation.images.top"]   # (3, 480, 640) float in [0, 1]
sample["observation.state"]        # (6,) degrees ×5 + gripper %
sample["action"]                   # (6,) leader-arm target, same units
sample["task"]                     # "pick up the x piece"

Held-out suggestion: the last episode of every task (20 episodes, one per task) — it keeps every language label represented in evaluation. Models trained on this split are published as cagataydev/ttt-act-so101 and cagataydev/ttt-smolvla-so101 (private while being evaluated).

Known limitations

• One scene, one operator, one lighting setup, one board position (± a few mm). No distractor objects, no occluded board.
• The piles are always on the same side; the O pieces are white on a white board (low contrast in the top view).
• Only 10 demonstrations per placing task — enough to fine-tune a pretrained VLA, thin for training from scratch per cell.
• The board is never pre-populated: no episode shows a piece being placed next to pieces already on the board, so a full game needs a policy that generalises to occupied boards, or a second recording pass.
• wrist is a true wrist camera here; in some other SO-101 datasets the slot named wrist is a fixed side view.

Citation

@misc{cagataycali2026so101ttt,
  title  = {SO-101 Tic-Tac-Toe: language-labelled pick-and-place demonstrations},
  author = {Çağatay Çalı},
  year   = {2026},
  url    = {https://huggingface.co/datasets/cagataydev/so101-tic-tac-toe}
}

Apache-2.0. Recorded for agi.diy / expo-the-arm. Card generated from meta/ by recording/push_card.py — the numbers above are read from the dataset, not typed.