rollout_eval_B_50ep

Real-robot evaluation rollouts of an ACT policy on the SO-101, red foam ball pick-and-place. Recorded 5 September 2026. This is evaluation data, not training data — every episode is the policy acting autonomously, no teleoperation.

Policy under test

Checkpoint	Jingyi-Z/act-sotac-runB-ep21-70 (step 100000)
Training data	Jingyi-Z/sotac episodes 21–70 (50 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

Sibling policies evaluated under the same protocol: Jingyi-Z/act-sotac-runA-ep0-70 (sotac ep 0–70, 71 episodes, rollouts in Jingyi-Z/rollout_eval_A_71ep) and Yu-Zhou-Wang/act_so_101_red_foam_ball (sotac ep 0–20, 21 episodes). A is the union of B and Y; B and Y are disjoint.

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 run 5 times in a row without moving anything — the ball is returned to its mark during the 30 s reset. The desk marks are identical to those used for rollout_eval_A_71ep (point 2 at its final position, point 7 included).

Bowl fixed at point 1, ball at each other point.

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, 26, 28, 29, 30	1-7	valid
—	27	(1-2)	discard — ball was placed on point 2 by mistake during the 1-7 block

31 episodes on disk, 30 valid. Episode 27 is kept for completeness but must be excluded from any score: the ball was on point 2 rather than point 7, so it belongs to neither scene. Episode 30 was recorded as a replacement so that scene 1-7 has five valid episodes. Point 2 is a known hard reach for the arm and is kept deliberately — all three policies face the same point.

Scoring

Scored offline from the recordings, not live, 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). Scores live in the SO101_eval_labeling_workbook_act_v2.xlsx workbook, not in this repo.

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("Jingyi-Z/rollout_eval_B_50ep")
valid = [e for e in range(ds.num_episodes) if e != 27]
scene_of = {**{e: "1-6" for e in range(0, 5)}, **{e: "1-5" for e in range(5, 10)},
            **{e: "1-4" for e in range(10, 15)}, **{e: "1-3" for e in range(15, 20)},
            **{e: "1-2" for e in range(20, 25)}, **{e: "1-7" for e in (25, 26, 28, 29, 30)}}

Notes

• Camera device indices on the recording machine were 0 (wrist) and 2 (top); indices 1 and 3 are the cameras' metadata nodes and are not capture devices.
• --display_data was off during recording: an X-forwarded Rerun viewer stalled the control loop to 3 Hz and froze the arm. Every frame is in the dataset regardless.