PranayTest/us-er2-policy-evals

US ER2 policy evals - ABC vs pi0.5 (piper-us-01, 2026-09-20)

Real-robot policy rollouts on the PiPER bimanual cell (plate pick, hand-over, place), run from the operator stack's Policy panel at execution speed multipliers of 1x to 6x. One folder per policy, one sub-folder per run (start time, IST): the three camera clips (top, left-arm, right-arm; re-encoded from the recorded chunks with every frame at its capture time, so they play in real time; clip t = 0 is 1 s before the run start and frame_ts.json lists each frame's capture time per camera), telemetry.jsonl / telemetry.parquet (every BigQuery signal in the window, including the policy's own policy_command_positions / policy_gripper_command) and manifest.json.

manifest.json carries the run parameters (params.speed is the time-scaling multiplier, 1.0 = the checkpoint's training speed), the gripper command events, the hand-over times (right close → left close → right release, seconds after the start), the peak commanded and measured joint speeds per arm and the stop reason. index.json is the same table as below. Telemetry shows commands and joint feedback, not whether the plate physically transferred.

Runs: 6; hand-over sequences completed: 3. Fastest complete hand-over: None at speed 2.0 on 2026-09-21T03:04:24.528000+05:30, right release at 66.8 s.

start (IST)	policy	speed	re-plan s	length s	R close	L close	R release	complete	R cmd/meas peak deg/s	stop
03:04:24	None	2.0	0.50	90.0	19.2	51.3	66.8	yes	109.6 / 72.1	duration reached
03:12:46	None	2.0	0.50	26.7	17.7	—	—	no	47.5 / 49.0	dashboard stop
03:14:25	None	2.0	0.50	137.7	7.8	82.2	99.0	yes	66.0 / 52.5	dashboard stop
05:16:25	None	1.0	0.50	150.1	37.3	—	—	no	59.6 / 46.9	duration reached
06:26:36	None	2.0	0.40	150.1	117.1	118.9	149.9	yes	59.0 / 49.3	duration reached
06:29:54	None	2.0	0.50	14.3	—	—	—	no	29.2 / 25.9	dashboard stop

Built by episode_pipeline/export_policy_runs.py from BigQuery + GCS; repo PranayTest/us-er2-policy-evals.