so101-tic-tac-toe-sim — SIMULATED twin of cagataydev/so101-tic-tac-toe

Used by cagataydev/agi-wfm-so101-ttt — the model that plays at api.agi.diy/play (65 % of moves completed in rendered sim, n = 180). Co-trained 1 : 1 by windows with the real set.

This is simulation, produced by a scripted IK expert (scripted-ik-ttt-v1), not teleoperation. It exists to be co-trained with the owner's real dataset cagataydev/so101-tic-tac-toe (210 real episodes, 20 tasks). Every episode carries meta/expert, meta/seed, meta/pace, meta/pitch_deg, meta/board_center, meta/board_yaw; meta/info.json carries agi_wfm.provenance (code commit d55bb59).

episodes / frames	2200 / 535456 (success-only)
fps	30
cameras	observation.images.top, observation.images.wrist, 480x640
state / action	6-D, names shoulder_pan.pos … gripper.pos, units so101-degrees-gripper-percent-v1 = LeRobot degrees for the arm, 0–100 % gripper — the same as the real dataset
tasks	the 20 real strings, verbatim (meta/tasks.parquet)
robot_type	so_follower
scene	MuJoCo (CPU), so101.xml (Menagerie) + Hashtag Robotics CAD board/tokens (CERN-OHL-P-2.0)

Co-trained into cagataydev/agi-wfm-so101-ttt, the flagship of api.agi.diy: 190 real + these 2,200 episodes, 1 : 1 by windows → 117 of 180 held-out rendered episodes solved (seed_base 7, disjoint from this set's seed_base 0). Rendered simulation only; the real-arm number does not exist yet. Collection: agi (model + the real dataset).

How the scene was matched to the real one (measured, agi_wfm/sim/ttt)

• Board + piles placed by FK of the 210 real episodes (cell lattice fit to 148 release points: 11.3 mm RMS vs the CAD 84 mm pitch); X pile robot-right, O pile robot-left.
• top camera fitted by PnP to the median real first frame: bar-crossing error mean 1.9 px / max 3.6 px. wrist camera fitted to the real wrist frames: jaw tips within 7 px.
• Expert copies the human: top-down grasp 3–16° from vertical, X by one cross arm, O by the rim (moving jaw in the hole), gripper opens only as far as the real 12–25 %; timing medians grasp 3.38 s / release 5.9 s / done 7.9 s vs real 3.7 / 6.3 / 8.5 s.
• Domain randomisation per episode: board pose ±1.5 cm / ±3°, table/sheet/bar/token colours, light direction+intensity, token poses + yaw, start pose drawn from the real rest-pose distribution.
• Known gaps (stated, not hidden): sim arm = ideal so101.xml kinematics — the real arm's calibration may carry a pan offset ≤10°; the real board moved between episodes; the real room/keyboard/cables are not modelled; sim lighting is one lamp.

Expert success at record time (success-only: failures were discarded)

task	kept	attempts	success
pick up the x piece	110	121	90.9 %
pick up the o piece	110	111	99.1 %
pick up the x piece and place it on the top left cell	110	129	85.3 %
pick up the x piece and place it on the top center cell	110	130	84.6 %
pick up the x piece and place it on the top right cell	110	128	85.9 %
pick up the x piece and place it on the middle left cell	110	129	85.3 %
pick up the x piece and place it on the center cell	110	125	88.0 %
pick up the x piece and place it on the middle right cell	110	130	84.6 %
pick up the x piece and place it on the bottom left cell	110	137	80.3 %
pick up the x piece and place it on the bottom center cell	110	134	82.1 %
pick up the x piece and place it on the bottom right cell	110	119	92.4 %
pick up the o piece and place it on the top left cell	110	112	98.2 %
pick up the o piece and place it on the top center cell	110	110	100.0 %
pick up the o piece and place it on the top right cell	110	113	97.3 %
pick up the o piece and place it on the middle left cell	110	112	98.2 %
pick up the o piece and place it on the center cell	110	110	100.0 %
pick up the o piece and place it on the middle right cell	110	111	99.1 %
pick up the o piece and place it on the bottom left cell	110	111	99.1 %
pick up the o piece and place it on the bottom center cell	110	110	100.0 %
pick up the o piece and place it on the bottom right cell	110	111	99.1 %

QA (agi_wfm.sim.verify_dataset on the merged bytes)

{
 "episodes": 2200,
 "frames": 535456,
 "windows": 465056,
 "ok": true,
 "problems": [],
 "follows_command": true,
 "mean_cosine_to_command": 0.773,
 "moved_toward_command_fraction": 0.917,
 "units": "so101-degrees-gripper-percent-v1"
}

Frame pts exact (max |pts − k/fps| = 0 ms on every file). Rows are (observation BEFORE the step, command OF the step).

Lane

EXPO-THE-ARM / TTT-SIM, 2026-10-26. Transfer measurement (done 2026-10-26 07:25Z, ACT 40k steps, real held-out 19,29,…,209, every 15th frame, degrees):

training data	3.333 s chunk RMSE	first-step	1.667 s chunk RMSE	tasks beating hold-last-action
hold-last-action (same windows)	25.42	—	18.18	—
sim-only (this dataset, 2,200)	28.30	9.95	23.25	5/20
real-only (190 real, cagataydev/ttt-act-so101)	17.86	3.88	12.48	19/20
mix 1:1 (190 real + 200 sim)	19.39	3.74	13.50	18/20
mix 1:3 (190 real + 580 sim)	17.73	3.53	12.82	19/20

A policy trained only on this dataset does NOT transfer to the real arm (worse than repeating the last action), and co-training with it did not improve a real-only ACT (1:3 level within 1 %, 1:1 worse). Artefacts: cagataydev/ttt-act-so101-{sim,mix1to1,mix1to3}; recipe repo cagataydev/so101-tic-tac-toe-mix. Use this dataset as a sim twin for tooling, scene/expert development and as a negative result — not as evidence of sim-to-real.

Successor: sim-v2

cagataydev/so101-tic-tac-toe-sim-v2 (4,104 episodes / 930,088 frames) extends this set with what a trained policy needed and this one does not contain: DAgger grasp recoveries recorded from the policy's own stalls, 18 clean-up strings (piece on a cell -> back to its pile), and heavier grasp-focused domain randomisation. Training on real + v1 + v2 lifted the grasp rate from 63.3 % to 71.1 % on the play board but scored 0/90 on clean-up — the v2 card records why.