so101-tic-tac-toe-sim-v2 — recovery + clean-up data for the SO-101 tic-tac-toe policy

Used by cagataydev/agi-wfm-so101-ttt — the model that plays at api.agi.diy/play, in its next post-tune. The served checkpoint was trained on real + sim v1 only; v2 is not in it yet.

Simulation (MuJoCo), not teleoperation. Second sim set for the real dataset cagataydev/so101-tic-tac-toe (210 teleop episodes, 20 strings) and its first sim twin cagataydev/so101-tic-tac-toe-sim (2,200 scripted-expert play episodes). v2 adds what the trained policy was measured to be missing — it does not repeat v1; train on real + v1 + v2.

Why this set exists (measured, not assumed)

The policy trained on real + v1 (agi-wfm/so101-ttt (plain-x45 @1000 = 4,000 effective it, served 2026-10-27), 65 % success on the n=180 rendered-sim board) fails almost only at the grasp — and rolling it in the sim 6,050 times showed how: it closes the gripper on nothing or slightly off (median max lift of the piece 5–6 mm), or lifts the piece and drops it (≈30 % of its stalls); it is at the wrong spot when it gives up (median 9 cm from the grasp point), not hovering patiently above the right one. v1's expert never visits those states (it retries from its own misses, not ours). And no dataset showed a piece leaving a cell, so the policy cannot clean the board for closed-loop play. v2 fixes exactly those two things:

section	episodes	frames	how
DAgger recovery	2587	538845	the served policy drove in sim; when it stalled the scripted expert took over from that state; only the recovery segment is written, under the ORIGINAL task string
clean-up	917	245678	18 new strings: the piece starts ON a cell and goes back to its pile (pick up the x piece from the center cell and place it back on the x pile)
play, grasp-heavy DR	600	145565	the 20 play strings under DR v2 (piles shifted ±2.5 cm, tokens crowded to 56 mm, desk brightness 0.30–0.66, wider piece colours and light)
total	4104	930088	success-only; 38 task strings

The DAgger funnel (the number this set is about)

Policy agi-wfm/so101-ttt (plain-x45 @1000 = 4,000 effective it, served 2026-10-27) (/work/ckpt_wfm in the jobs), 6050 rollouts on the play strings: 2422 succeeded on their own (40 %), 3628 stalled and were handed to the expert, 2587 recoveries kept (628 the expert could not finish). Hand-over reasons: {"closed_no_lift": 1726, "no_lift": 750, "policy_success": 2422, "cap": 32, "dropped": 1120}; median hand-over at 4.42 s. Stall = no token of the named piece above 3 cm by a per-episode 4–8 s, or gripper closed with nothing lifted (1 s in recorder v1, 2.5 s = one chunk in v2), or a lifted piece dropped away from the target, or 12 s. The 20 s board gives this policy 65 %; the early hand-over is why "succeeded on their own" reads lower here — the cut rollouts become recoveries from mid-approach states, which is intended.

Recorders: {"dagger-recovery-v1": 1387, "dagger-recovery-v2": 1200} episodes. Hand-over modes (v2 recorder): {"regrasp": 1028, "continue_lift": 172} — continue_lift = the policy had the piece, the expert finished the lift; regrasp = open, re-approach. 413 recoveries from the v1 recorder were dropped (closed-gripper hand-overs with the piece already >= 10 mm up: the policy was lifting slowly and a re-grasp label would have contradicted it). The per-task rows below count kept episodes BEFORE that filter.

task	rollouts	policy succeeded	recoveries kept (pre-filter)
pick up the x piece	253	91	150
pick up the o piece	450	285	150
pick up the x piece and place it on the top left cell	218	23	150
pick up the x piece and place it on the top center cell	249	66	150
pick up the x piece and place it on the top right cell	255	57	150
pick up the x piece and place it on the middle left cell	245	59	150
pick up the x piece and place it on the center cell	273	86	150
pick up the x piece and place it on the middle right cell	273	76	150
pick up the x piece and place it on the bottom left cell	222	34	150
pick up the x piece and place it on the bottom center cell	280	74	150
pick up the x piece and place it on the bottom right cell	275	97	150
pick up the o piece and place it on the top left cell	268	81	150
pick up the o piece and place it on the top center cell	342	157	150
pick up the o piece and place it on the top right cell	369	180	150
pick up the o piece and place it on the middle left cell	435	252	150
pick up the o piece and place it on the center cell	374	199	150
pick up the o piece and place it on the middle right cell	261	95	150
pick up the o piece and place it on the bottom left cell	316	147	150
pick up the o piece and place it on the bottom center cell	317	155	150
pick up the o piece and place it on the bottom right cell	375	208	150

Format

fps	30
cameras	observation.images.top, observation.images.wrist, 480x640 (the real top / wrist names)
state / action	6-D shoulder_pan.pos … gripper.pos, units so101-degrees-gripper-percent-v1 = LeRobot degrees for the arm, 0–100 % gripper — identical to the real dataset and v1
robot_type	so_follower
per-episode meta	meta/expert, meta/seed, meta/dr_profile, meta/source_part; clean-up: meta/start_cell, meta/pile_slot; recovery: meta/recorder, meta/dagger_policy, meta/handover_s, meta/handover_reason, meta/policy_max_lift_m, meta/policy_calls, meta/policy_grip_closed
scene	MuJoCo (CPU/EGL), so101.xml (Menagerie) + Hashtag Robotics tic-tac-toe board and tokens, CAD under CERN-OHL-P-2.0 (credit: Hashtag Robotics)
seeds	v1 dataset 0 · success boards 7 · clean-up 2 · play DR v2 3 · DAgger 11 — disjoint, no evaluated layout is a training layout
code	agi-wfm a4defe9 (agi_wfm/sim/ttt: scene.py, expert.py, record.py, dagger.py, publish_v2.py)

Rows are (observation BEFORE the step, command OF the step), 30 Hz, like a teleop recording. A recovery episode starts mid-air — gripper at the piece, open or wrongly closed — exactly the state the owner's recording plan 2 (section C) reproduces on the real arm.

Expert success at record time (failures discarded)

Clean-up (DR v2):

task	kept	attempts	success
pick up the x piece from the top left cell and place it back on the x pile	60	95	63 %
pick up the x piece from the top center cell and place it back on the x pile	60	73	82 %
pick up the x piece from the top right cell and place it back on the x pile	60	75	80 %
pick up the x piece from the middle left cell and place it back on the x pile	60	73	82 %
pick up the x piece from the center cell and place it back on the x pile	60	76	79 %
pick up the x piece from the middle right cell and place it back on the x pile	60	74	81 %
pick up the x piece from the bottom left cell and place it back on the x pile	60	80	75 %
pick up the x piece from the bottom center cell and place it back on the x pile	60	91	66 %
pick up the x piece from the bottom right cell and place it back on the x pile	60	100	60 %
pick up the o piece from the top left cell and place it back on the o pile	6	100	6 %
pick up the o piece from the top center cell and place it back on the o pile	34	100	34 %
pick up the o piece from the top right cell and place it back on the o pile	33	100	33 %
pick up the o piece from the middle left cell and place it back on the o pile	21	100	21 %
pick up the o piece from the center cell and place it back on the o pile	47	100	47 %
pick up the o piece from the middle right cell and place it back on the o pile	56	100	56 %
pick up the o piece from the bottom left cell and place it back on the o pile	60	60	100 %
pick up the o piece from the bottom center cell and place it back on the o pile	60	60	100 %
pick up the o piece from the bottom right cell and place it back on the o pile	60	60	100 %

Play (DR v2):

task	kept	attempts	success
pick up the x piece	30	33	91 %
pick up the o piece	30	30	100 %
pick up the x piece and place it on the top left cell	30	37	81 %
pick up the x piece and place it on the top center cell	30	33	91 %
pick up the x piece and place it on the top right cell	30	35	86 %
pick up the x piece and place it on the middle left cell	30	35	86 %
pick up the x piece and place it on the center cell	30	31	97 %
pick up the x piece and place it on the middle right cell	30	36	83 %
pick up the x piece and place it on the bottom left cell	30	36	83 %
pick up the x piece and place it on the bottom center cell	30	38	79 %
pick up the x piece and place it on the bottom right cell	30	35	86 %
pick up the o piece and place it on the top left cell	30	30	100 %
pick up the o piece and place it on the top center cell	30	30	100 %
pick up the o piece and place it on the top right cell	30	30	100 %
pick up the o piece and place it on the middle left cell	30	30	100 %
pick up the o piece and place it on the center cell	30	30	100 %
pick up the o piece and place it on the middle right cell	30	30	100 %
pick up the o piece and place it on the bottom left cell	30	30	100 %
pick up the o piece and place it on the bottom center cell	30	30	100 %
pick up the o piece and place it on the bottom right cell	30	31	97 %

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

{
 "episodes": 4104,
 "frames": 930088,
 "windows": 798760,
 "ok": true,
 "problems": [],
 "follows_command": true,
 "mean_cosine_to_command": 0.757,
 "moved_toward_command_fraction": 0.909,
 "units": "so101-degrees-gripper-percent-v1"
}

Honest limits

Rendered frames — the camera fits are v1's (top 1.9 px, wrist 7 px); nothing here is measured on the real arm. The recovery states are the sim policy's stall states; the owner's real recovery recordings (plan 2) are the real-arm counterpart. Judge for everything downstream: ttt-sim-judge-v1 (agi_wfm/sim/ttt/closed_loop.py).

Lane TTT-DATA2, 2026-10-27.

What training on this data produced (2026-10-27)

The first post-tune to use v2 — cagataydev/agi-wfm-so101-ttt's candidate mix-v2-x45, real 1 : sim-v1 0.5 : sim-v2 0.5 by windows, warm-started from the served checkpoint, 1,500 iterations — was measured against gates fixed in advance and was not promoted:

• Grasp: better. On the 180-episode play board it lifted the piece in 71.1 % of episodes against the served checkpoint's 63.3 %, and success held at 120/180 (66.7 %) vs 117/180 — unchanged within noise. The DAgger half of this dataset did what it was built for.
• Clean-up: no transfer at all. 0/90 on the 18 clean-up strings. Not a near miss: the median episode never lifts the piece (0.2 mm) and ends 20 cm from the pile, the closest carry was 4.6 cm against a 4.2 cm tolerance, and the policy knocks other tokens in 78 of 90 episodes where the scripted expert does so in 6. The scripted expert reaches 67/90 on the same board, so the ceiling is not the problem.
• Real frames: unharmed. Held-out RMSE 13.36 over 2.13 s, the best of any checkpoint in the program (served 13.71), so adding 930k rendered frames did not degrade the fit to the owner's real recordings.

If you train on the clean-up half, change two things first. Clean-up was only 6.3 % of all training windows (26.4 % of this dataset's frames x 23.8 % realised share), and the 18 clean-up strings share almost their whole grammar with the 20 play strings ("pick up the x piece from the center cell and place it back on the x pile" vs "pick up the x piece and place it on the center cell"). The failure looks like string confusion rather than motor incapacity — the same checkpoint grasps fine on play strings — so re-weight the clean-up windows and make the strings lexically distinct. That hypothesis is consistent with the three numbers above but has NOT been tested; the cheap test is a task-condition probe on one clean-up observation against the 9 play strings.