Datasetpublic

Jingyi-Z/rollout_eval_tac_mlp

rollout_eval_tac_mlp Rollout evaluation of Jingyi-Z/act-sotac-tac-mlp (ACT + raw tactile frame token, trained on sotac ep 0-70) on the SO-101 bench, 2026-09-16. Protocol: bowl fixed at point 1, ball at points 6, 5, 4, 3, 2, 7; five consecutive 60 s episodes per scene, 30 s reset. Stock lerobot 0.6.1 rollout with the so101_paxini robot type; observation.state is 318-dim (6 joints + 312 tactile forces at 30 Hz), so tactile is recorded in every episode. Cameras: top idx 0, wrist… See the full description on the dataset page: https://huggingface.co/datasets/Jingyi-Z/rollout_eval_tac_mlp.

sourceHugging Faceupdated 8d agoView on Hugging Face

0likes89downloads

Dataset Card

rolloutevaltac_mlp

Rollout evaluation of Jingyi-Z/act-sotac-tac-mlp (ACT + raw tactile frame token, trained on sotac ep 0-70) on the SO-101 bench, 2026-09-16.

Protocol: bowl fixed at point 1, ball at points 6, 5, 4, 3, 2, 7; five consecutive 60 s episodes per scene, 30 s reset. Stock lerobot 0.6.1 rollout with the so101_paxini robot type; observation.state is 318-dim (6 joints + 312 tactile forces at 30 Hz), so tactile is recorded in every episode. Cameras: top idx 0, wrist idx 2.

Episode map

Episodes	Scene	Note
0-4	discard	tactile finger order swapped vs training
5-7	discard	stale calibration, systematic reach offset
8-12	1-6	valid
13-17	1-5	valid
18-22	1-4	valid
23-27	1-3	valid
28-32	1-2	valid
33-37	1-7	valid

Valid episodes: 8-37 (30 total). Exclude 0-7 by index when scoring.

Eval-time config notes: fingertip slots mapped to match sotac's finger order (moving jaw first) via a processor-step flag; arm calibration = the bench file current at recording time. High-rate raw sensor sessions were recorded on the bench alongside but are not part of this dataset.

Scoring: stage rubric S0-S5 in the lerobotac dataset visualizer, as for the baseline evals. Companions: rolloutevaltac{pred,esn,esnpred}, baselines rolloutevalA71ep / B / Y.