rollout_eval_A_sd50

Real-robot evaluation rollouts of a state-dropout ACT policy on the SO-101, red foam ball pick-and-place. Recorded 11 September 2026, two bowl positions in one session. This is evaluation data, not training data — every episode is the policy acting autonomously.

Policy under test

Checkpoint	Jingyi-Z/act-sotac-runA-sd50 (step 100000)
Training data	Jingyi-Z/sotac episodes 0–70 (71 episodes), tactile column removed
Inputs	observation.state (6) + observation.images.top + observation.images.wrist — no tactile
Training	Same as act-sotac-runA-ep0-70 plus state dropout 0.5: each training sample's proprio vector zeroed with probability 0.5. 100k steps, batch 8, seed 1000, lr 1e-5, ResNet18, chunk 100, no image augmentation, MEAN_STD normalisation. Final loss 0.043, identical to the no-dropout baseline

Direct comparison target: Jingyi-Z/rollout_eval_A_71ep (same training data, no state dropout, same scene set). At rollout, state_dropout_prob was stripped from the downloaded config.json so the older lerobot could load it; the masking is training-only.

Rollout setup

Robot	SO-101 follower, calibration id so_sensor_follower
Cameras	wrist = Innomaker U20CAM-1080p (/dev/video2), top = icspring (/dev/video0); both 640×480 @ 30 fps
Control	30 fps, sync inference, lerobot-rollout --strategy.type=episodic
Episode	60 s policy execution, 30 s reset between episodes
Task string	"Pick up the red foam ball and place it into the bowl"
Recorder	lerobot (a0eval env), push_to_hub=false, uploaded afterwards

Scene design

Seven marked points on the desk, identical to the sibling evaluations. A scene is a (bowl point, ball point) pair, written bowl-ball, 5 consecutive episodes per scene.

Scene	Episodes	Bowl–Ball	Status
1	0–4	1-6	valid
2	5–9	1-5	valid
3	10–14	1-4	valid
4	15–19	1-3	valid
5	20–24	1-2	valid
6	25–29	1-7	valid
7	30–34	6-1	valid
8	35–39	6-2	valid
9	40–44	6-3	valid
10	45–49	6-5	valid
11	50–54	6-7	valid

Scene 6-4 was skipped: the bowl at point 6 physically covers point 4 (about 5 cm apart). 55 episodes on disk, 55 valid, no discards.

Scoring

Scored offline from the recordings using the lab rubric: score = highest stage reached, not additive — S0 freeze 0 · S1 reach 0.2 · S2 grasp 0.4 · S3 transport 0.7 · S4 release 0.8 · S5 ball in bowl 1.0 — plus one primary failure mode (F0_none … F8_other).

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("Jingyi-Z/rollout_eval_A_sd50")
scenes = ["1-6", "1-5", "1-4", "1-3", "1-2", "1-7", "6-1", "6-2", "6-3", "6-5", "6-7"]
scene_of = {e: scenes[e // 5] for e in range(55)}