Kavin60606/so101-ma2-rollouts

SO-101 pick_marker — MolmoAct2 fine-tunes: real rollouts vs CI-MSE

Three checkpoints of one fine-tuning recipe (Kavin60606/so101-pick-marker-ma2-ckpts: s2000/001750, s1000/001750, s1000/002000) evaluated two ways:

• Offline, CI-MSE exactly as in the CI-MSE paper/repo: held-out episodes 45–49 (2 276 frames), open-loop 30-step action-chunk predictions, critical intervals labelled by a VLM (gemini-2.5-pro, CI-MSE prompt template, 3 samples per episode, median), DTW alignment (interpolation 5), no temporal ensemble, raw action units — i.e. python -m val_metrics.validate dump.h5 --intervals ci_vlm.json with default flags. Whole-episode MSE = --no-critical-intervals --no-dtw.
• Online, real robot: SO-101 follower + wrist camera on a Mac, policy served on a remote GPU, 20 rollouts per checkpoint (10 box-vertical, 10 box-horizontal, same marker mark), 30 s timeout, success = marker in box and gripper released. Stages graded reach → grasp → carry → place.

Results (2026-09-30 / 10-01)

ckpt	CI-MSE (VLM intervals + DTW)	rank	whole-episode MSE	rank	real SR	n	Wilson 95 %	ρ CI-MSE↔SR	r	pairs	ρ MSE↔SR	r	pairs
s2000/001750	84.8	1	110.7	1	75 %	20	[53, 89]	-1.00	-0.99	3/3	-1.00	-0.99	3/3
s1000/001750	89.4	2	113.1	2	60 %	20	[39, 78]
s1000/002000	97.6	3	116.9	3	45 %	20	[26, 66]

ρ = Spearman rank correlation across the three checkpoints, r = Pearson; metrics are errors (lower is better) so −1 is perfect agreement with success rate.

ckpt	reach	grasp	carry	box vertical	box horizontal	main failure
s2000/001750	95 %	85 %	80 %	9/10	6/10	mixed
s1000/001750	65 %	60 %	60 %	6/10	6/10	no-reach
s1000/002000	60 %	60 %	55 %	4/10	5/10	no-reach

Pairwise two-proportion tests (20 vs 20): rank 1 vs rank 3 z = 1.94 (p ≈ 0.05); the two 15-point gaps p ≈ 0.3.

Caveats

• n = 3 checkpoints: ρ = −1 is the best attainable value (1-in-6 by chance) and does not separate CI-MSE from whole-episode MSE on this data.
• Same operator resets and grades; conditions blocked (10 V then 10 H), not interleaved.
• Deployment identical for all checkpoints: training joint-range clip ±5°, 5°/step velocity clamp, servo torque limits, scripted 0.5 s onset kick, async chunk execution.
• Flow-matching sampling in bf16 is not bit-reproducible across GPUs (~3 % run-to-run movement in the offline metrics).

Files

rollouts.csv (one row per rollout), cimse_vlm/ — ci_vlm.json (intervals, frames @ 30 fps), ci_vlm_raw.json (all VLM samples), PickPlaceMarker.json (prompt), annotate_or.py (annotator), vlm_intervals_check.png (interval sanity frames), table_3ckpts.json (all metric values); prediction dumps for re-scoring live in the checkpoint repo under real_rollouts/cimse_vlm/dumps/. Per-rollout folders <ckpt>/<V|H>_repNN_<timestamp>/{trace.npz, wrist.mp4}.