ghkim-rlwrld/race-realworld

RACE — real-robot planner data

Two tasks on one 7-DoF arm, chosen to sit on different axes. Both give a high-level planner the same contract: read the scene, decide the subtask the controller should execute next, and say whether the accepted one has finished.

shell_game        memory     the answer is only in frames that have already scrolled past
solve_equation    reasoning  the answer is in view but has to be computed

Same rig for both: 7-DoF arm plus gripper, three stereo cameras (11845904 wrist, 37149196 ext_left, 34022131 ext_right), 9.2 Hz measured control rate — the mp4 containers claim 10 and are wrong.

Layout

shell_game/
  rows.jsonl              1,171 planner calls over 45 episodes
  frames/                 2,777 wrist frames the rows reference
  scenarios.json          the answer key: answerSide per episode, and swaps as the difficulty axis
  subtask_segments.jsonl  state-derived boundaries, agreeing with the collection sheet on 44/45

solve_equation/
  rows.jsonl              492 planner calls over 20 episodes
  frames/                 wrist frames
  subtasks.jsonl          60 placements: which staging slot went into which gap, and when

What a row is

A system message carrying the contract, a user message carrying the goal and the subtask the controller last accepted, an ordered image list, and the target: a 1/0 transition token, a newline, then {"current_subtask": ..., "keyframe_positions": [...]}.

Images are memory first, then the recent window, and both are built by the rules the deployed client uses, so every row is a prompt the client can actually produce. That property was not free — an earlier build put shell_game's shuffle in memory at a stride the client never uses, and the rows were unreachable at deployment.

	shell_game	solve_equation
camera	wrist	wrist
memory	frames committed at past transitions, cap 9	frames 0 and 20 pinned, plus completed placements
recent	from the episode start, stride 11, cap 21	stride 8, 3 frames
images/row	median 21	median 6
transition rows	7.7 %	8.1 %

The two differ because the information does. shell_game's evidence is spread over a shuffle where one swap spans ~24 frames and there are up to five, so a stride of 24 samples about one frame per swap and the shuffle cannot be recovered; 11 guarantees two. solve_equation's evidence is a printed sheet that does not change, so one pinned frame carries it — with a second pinned frame because the gripper covers a staging block in 3 of 20 episodes at frame 0 and clears by frame 20.

Scoring

shell_game. scenarios.json gives answerSide, the cup the block ends under. The floor is not zero: always answering "left" scores 40.0 % and lifting the cup the block started under scores 33.3 %. swaps (1-5, at 5/8/10/11/11 episodes) is the difficulty axis, and the slope against it separates a model that tracks from one that guesses better than the headline number.

solve_equation. Digits and target are printed, and exactly one operator arrangement satisfies them, so the answer can be derived rather than looked up — 18 of 20 episodes have a unique solution over all 27 triples. subtasks.jsonl records which staging slot went into which gap. Gaps fill left to right in every episode, so the gap is a function of the step index and the live choice is which of four blocks to take.

What is already known

shell_game, zero-shot on the same 45 episodes, asked as a plain question:

	grounding	shuffle tracking
Qwen3-VL-4B	95.6 %	35.6 %
Qwen3-VL-8B-Instruct	100.0 %	24.4 %
chance	33.3 %	33.3 %

grounding asks which cup the arm is lifting from frames where it already is. Perception is solved and tracking through occlusion is at chance, and the larger model is perfect at the first and worse at the second.

solve_equation, Qwen3-VL-4B, all three operators in one answer:

  48 tok, answer only     0 %          2048 tok   28 %
 512 tok                  0 %          4096 tok   50 %
1024 tok                  6 %          8192 tok   61 %

Every finished answer was fully correct and there were no partial ones: the model does not get it wrong, it runs out of room. Immediate answering with the reply prefilled is 0/18 at four digits, 47 % at three and 100 % at two.

Caveats

Every episode is in these sets; there is no held-out split. n = 45 and n = 20. The labels are derived from cartesian y at the grasp and release instants and were never read off the video, so the rollout videos are the independent check on them.