act-so101-pickplace-dr2 — targeted-coverage retrain (a coverage whack-a-mole result)

ACT (~52M) trained from scratch on so101-sim-pickplace-dr2 — 210 domain-randomized SO-ARM100 pick-and-place demonstrations of which ~30% deliberately oversample a small near-field spawn band (r∈[0.19,0.23], θ∈[−14°,+2°]) covering the pocket where the previous best policy, ACT-DR, put all 5 of its 50-episode-eval failures (spawn r ≈ 0.19–0.21 m). Same recipe otherwise (chunk 45, n_action_steps 15, batch 8; 75k steps to approximately match epochs on the larger set, ~25.7 vs ~28.7). The hypothesis was ≥95% nominal. The result is a documented negative: the targeted data closed the measured hole everywhere we probed — and opened a new one.

Results

50 evaluation episodes, fixed eval seed (identical spawn sequence for both models):

The two failure sets share only episode 26. dr2 fixed 4 of ACT-DR's 5 failures and introduced 5 new ones — a coverage swap, confirmed on fixed spawn grids (scripts/probe_spawns.py):

Robustness sweep (15 eps/level, same seed across levels):

Temporal ensembling (inference-only, coeff 0.01, 20 eps): 18/20 (90%) — comparable to ACT-DR's 18/20 under the same protocol.

What this negative result teaches

• Oversampling a failure pocket inside a fixed step budget appears to reallocate precision rather than add it. Under an equal-weight L1 loss, ~30% of gradient mass now serves a narrow near-field band; per-epoch relative exposure to the rest of the workspace drops correspondingly, and the policy regressed exactly where demands are highest — near maximum reach (r ≥ 0.259 m, with the spawn range capped at 0.27 m). ACT-DR, trained on the uniform 150-episode set, is 12/12 there. (Caveat: dr2 also differs in episode count and steps; the isolating control is below.)
• Point evals hide this. Nominal success barely moved (90% → 88%, well within noise at n=50). Only the per-episode failure comparison and the fixed spawn grids exposed that the model is different, not merely equal.
• The residual shared failure (episode 26, the nearest-to-base spawn) resists both data distributions.

Follow-ups this predicts: append pocket episodes to the full uniform set at a lower mixing fraction (e.g. 10–15%), or reweight per-sample loss instead of resampling — either preserves far-field exposure while still covering the pocket. The cleanest attribution experiment is an untargeted 210-episode control retrain at 75k steps, which isolates the reweighting effect from the episode-count/step changes.

Training

• LeRobot 0.6.0 lerobot-train, single Kaggle T4, AMP — 75k steps (~25.7 epochs of 210 episodes; ACT-DR: 60k steps ≈ 28.7 epochs of 150)
• --policy.type=act --policy.chunk_size=45 --policy.n_action_steps=15 --batch_size=8

Use

from lerobot.policies import make_pre_post_processors
from lerobot.policies.act.modeling_act import ACTPolicy

repo = "ahmedsohail2003/act-so101-pickplace-dr2"
policy = ACTPolicy.from_pretrained(repo).to("cuda").eval()
pre, post = make_pre_post_processors(
    policy_cfg=policy.config, pretrained_path=repo,
    preprocessor_overrides={"device_processor": {"device": "cuda"}},
)
# batch: observation.state (1,6) + observation.images.front / .wrist (1,3,224,224) in [0,1]
action = post(policy.select_action(pre(batch)))

The Sim2Cell ACT family

Full study in the Part A README.

Provenance

Data recorded by a scripted expert (success-filtered) in the author's MuJoCo work-cell; robot model from MuJoCo Menagerie (trs_so_arm100, Apache-2.0).