Sa74ll/LocalModels-ACT-bimanual-rollouts

LocalModels / ACT-bimanual: GRPO rollouts

Rollouts saved by team LocalModels during the CoRL 2026 RoboEval challenge: 29 GRPO fine-tuning runs of ACT policies on six bimanual tasks, 389,856 episodes (318,256 successes, 75.8 million steps, 3.96 GB). Each episode is stored as its environment seed plus every executed action, about 10 kB instead of about 40 MB rendered. With pinned BLAS the simulator is deterministic, so an episode replays exactly and its camera frames can be re-rendered.

The final entry ships one learned model, LiftPot: LiftPot/grpo_02 is the GRPO stage that produced it. The other runs (including LiftPot/grpo_01) fine-tuned ACT models that the final entry does not use.

• Models and the official run: Sa74ll/LocalModels-ACT-bimanual
• Code and report: github.com/Sa74ll/RoboevalChallenge

Runs

Runs are numbered per task in the order of their GRPO round numbers. Rollouts use training seeds (10,000 and above): 8 noisy rollouts per start state, Gaussian noise (sigma 0.2 or 0.3) on the normalized actions.

Task	Run	Rounds	Episodes	Success	Notes
LiftPot	grpo_01	31000-31116	11,232	75%	restart from GRPO v1 round 1150; was lp_l1
LiftPot	grpo_02	37000-37199	19,200	84%	trained the shipped LiftPot model (snapshot round 37100: rounds 37000-37099, 9,600 episodes); was lp_l2
PackBox	grpo_01	18045-18199	14,880	81%	was gpb9500_t1
PackBox	grpo_02	20000-20086	8,352	78%	was pb9500b_v2_5090
PackBox	grpo_03	30000-30133	12,864	84%	was gpb_p1
PackBox	grpo_04	41000-41199	19,200	82%	task-progress reward bonus; was pbtp_5090
PackBox	grpo_05	43000-43199	19,200	91%	was pb_p3_5090
PackBox	grpo_06	44000-44261	16,768	90%	was pb_p4_5090
PackBox	grpo_07	48000-48499	32,000	92%	was pb_p5_5090
PackBox	grpo_08	49000-49249	24,000	94%	was pb_p6
PackBox	grpo_09	50000-50316	30,432	94%	was pb_p7_5090
PickSingleBookFromTable	grpo_01	34000-34199	19,200	82%	was psb_k1
PickSingleBookFromTable	grpo_02	46000-46156	10,048	83%	was psb_k2
RotateValve	grpo_01	45000-45089	5,760	86%	was rv_r2
RotateValve	grpo_02	47000-47299	28,800	88%	was rv_r3
StackSingleBookShelf	grpo_01	23000-23024	2,400	61%	action scale 0.85; was ssbs160_v2
StackSingleBookShelf	grpo_02	28000-28009	960	76%	action scale 0.85; was ssbsm1_v2
StackSingleBookShelf	grpo_03	35000-35199	19,200	88%	action scale 0.85; was ssbsm2_v2
StackSingleBookShelf	grpo_04	36000-36199	19,200	90%	action scale 0.85; was ssbsm2_v2b
StackTwoBlocks	grpo_01	16045-16199	14,880	58%	was s2bbr_v2b
StackTwoBlocks	grpo_02	17060-17199	8,960	60%	was s2bbr_a3
StackTwoBlocks	grpo_03	21075-21149	7,200	58%	was s2bbr_v2b_f5090
StackTwoBlocks	grpo_04	22075-22174	9,600	60%	was s2bbr_a4
StackTwoBlocks	grpo_05	25075-25086	1,152	59%	privileged value baseline; was s2bbr_a5
StackTwoBlocks	grpo_06	26000-26161	15,552	63%	privileged value baseline; was s2bbr_a5b
StackTwoBlocks	grpo_07	27000-27079	7,680	53%	was s2bbr_a6
StackTwoBlocks	grpo_08	29000-29054	5,280	54%	was s2bbr_a6b
StackTwoBlocks	grpo_09	32000-32057	5,568	57%	was s2bbr_a7
StackTwoBlocks	grpo_10	33000-33002	288	61%	was s2b_a8

Notes: ShelfBook runs executed the policy's actions scaled by 0.85 (the stored actions are the executed ones). All StackTwoBlocks rewards include shaping terms computed from simulator positions, and two of its runs used a value baseline from simulator state; the stored data are still seeds and actions only.

Format

<Task>/<run>/log.jsonl                one line of GRPO metrics per round
<Task>/<run>/successes/it<N>.npz      round N, episodes that succeeded (+ it<N>.json)
<Task>/<run>/failures/it<N>.npz       round N, all other episodes      (+ it<N>.json)

• it<N>.npz: actions (float32, steps x 14: end-effector deltas dp_L dr_L dp_R dr_R g_L g_R at 20 Hz) and offsets (episode k is actions[offsets[k]:offsets[k+1]]).
• it<N>.json: one record per episode: seed, steps, term (how it ended), success, tp (task progression), R (GRPO reward), iter, prefix_t (actions replayed from a logged episode before the rollout; 0 for a fresh start).

To replay an episode: env.reset(seed=seed), then step every stored action (the first is a one-step open-gripper no-op). Exact replay needs RoboEval at tag corl2026-challenge, MuJoCo 3.3.3 and pinned BLAS (OPENBLAS_CORETYPE=Haswell OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 MKL_CBWR=COMPATIBLE).

License

Apache-2.0. No demonstration data or rendered frames are included.