rollout_eval_A_71ep

Real-robot evaluation rollouts of an ACT policy on the SO-101, red foam ball pick-and-place. Two sessions: 5 September 2026 (bowl at point 1) and 11 September 2026 (bowl at point 6). This is evaluation data, not training data — every episode is the policy acting autonomously, no teleoperation.

Policy under test

Checkpoint	Jingyi-Z/act-sotac-runA-ep0-70 (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	100k steps, batch 8, seed 1000, lr 1e-5, ResNet18, chunk 100, no augmentation, MEAN_STD normalisation

Siblings under the same protocol: Jingyi-Z/rollout_eval_B_50ep, rollout_eval_Y_21ep, rollout_eval_A_sd50 (state-dropout variant), plus aborted runs rollout_eval_DP_71ep and rollout_eval_SV_71ep.

Rollout setup

Robot	SO-101 follower, calibration id so_sensor_follower
Cameras	wrist = Innomaker U20CAM-1080p, top = icspring; 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. A scene is a (bowl point, ball point) pair, written bowl-ball. Each scene is 5 consecutive episodes, ball returned to its mark in the reset.

Session 1 (5 Sep): bowl at point 1

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
—	20–24	1-2	discard — original point 2, out of reach, marker moved
5	25–29	1-2	valid — point 2 at its final position
6	30–34	1-7	valid — point 7 added
—	35–36	1-8	discard — point 8 abandoned after 2 episodes

Session 2 (11 Sep): bowl at point 6

Scene	Episodes	Bowl–Ball	Status
7	37–41	6-1	valid
8	42–46	6-2	valid
9	47–51	6-3	valid
10	52–56	6-5	valid
11	57–61	6-7	valid

Scene 6-4 was skipped: points 6 and 4 are about 5 cm apart, so the bowl at point 6 physically covers point 4. Camera device indices were swapped in session 2 after a USB replug (wrist on /dev/video2, top on /dev/video0); the observation keys are correct in both sessions.

62 episodes on disk, 55 valid (30 from session 1, 25 from session 2). Discards stay in the dataset and are excluded by index.

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_71ep")
discard = {*range(20, 25), 35, 36}
s1 = ["1-6"]*5 + ["1-5"]*5 + ["1-4"]*5 + ["1-3"]*5 + ["x"]*5 + ["1-2"]*5 + ["1-7"]*5 + ["x"]*2
s2 = ["6-1"]*5 + ["6-2"]*5 + ["6-3"]*5 + ["6-5"]*5 + ["6-7"]*5
scene_of = {e: s for e, s in enumerate(s1 + s2) if s != "x"}

Notes

• --display_data was off during recording: an X-forwarded Rerun viewer stalled the control loop and froze the arm. Every frame is in the dataset regardless.