Kavin60606/cimse-so100-experiment-v2

⚠️ ERRATUM (2026-09-13). The so100_bowl_* datasets used for this experiment had misaligned language labels: the ManiSkill→LeRobot converter stored episodes in lexicographic key order while per-episode task strings were assigned in numeric order, so ~2/3 of training and validation episodes named the wrong cube colour. Consequences for the results below: (1) the 21 checkpoints were trained on effectively random colour labels and never learned grounding — the 30–60 % "wrong cube" failure rate is this bug, not a policy property; (2) the sim-state GT critical intervals were indexed in raw order and are misaligned with the prediction dumps (the Gemini/VLM intervals were annotated from the dataset videos and are aligned); (3) offline dumps used the same wrong task strings as training. The metric-vs-success correlations are internally consistent for an ungrounded policy, but every conclusion about language, phase attribution and the data-scale family must be treated as void until the zoo is retrained on the relabelled datasets (labels fixed on HF on 2026-09-13). v1 (cimse-so100-experiment, single task, no language) is unaffected.

Summary

We asked whether an offline metric computed on a frozen validation set can rank policy checkpoints the way closed-loop success does, in the regime the CI-MSE paper targets: noisy, multimodal demonstrations and a language-conditioned task. One policy family (MolmoAct2, 5B), 21 checkpoints, 200 closed-loop episodes each, full-frame offline dumps on 282 validation episodes, three sources of critical-interval annotation.

Findings (21 checkpoints, open-loop success 0.00–0.28):

1. Plain action MSE over all validation frames ranks checkpoints about as well as anything we tried: Spearman ρ = −0.80 (95 % bootstrap CI [−0.86, −0.56]), zero selection regret. Same conclusion as v1.
2. CI-MSE without ensembling matches it (ρ = −0.73 with GT windows, −0.77 with Gemini median-of-3 windows). Restricting to critical intervals neither helps nor hurts here, largely because the windows cover ~95 % of frames.
3. The ensembled CI-MSE variant is clearly worse (ρ = −0.42 GT / −0.55 VLM; CI [−0.55, −0.19]) because it scores a controller the robot never runs. Deployed with that controller (temporal ensembling), every checkpoint collapses to 0–10 % success. The offline metric must mirror the deployed controller; the paper's ensembling assumption is not policy-agnostic.
4. Gemini-annotated windows are as good as sim-state ground truth, and median-of-3 sampling is slightly better than a single sample. Annotation quality is not the bottleneck.
5. Phase attribution does not work: window-specific errors do not isolate phase-specific failures. The dominant failure (wrong cube, 18–59 %) rises with training because it is conditional on grasping at all.
6. Data scale is not resolved by any metric (5 % data @3000 is the best checkpoint at 0.28; 100 % is 0.26; 40 % is 0.16). With 200 episodes these differ within noise and no offline metric orders them.

1. Setup

• Task: "put the {red|green|blue} cube in the bowl". ManiSkill3 GPU sim, SO-100, 30 Hz, two 224² cameras, three cubes (1.25–1.45 cm), a bowl, domain randomisation (lighting, camera mounts ±2 cm/±3°, fov, colour), 450-step cap.
• Demonstrations: 2812 training episodes from three scripted oracles — S0 clean (1146), S1 noisy with pauses, misses and regrasps (859), S2 strategy-varied: speed, carry height, yaw-90 grasps, lateral approaches (807). Validation: s0 97 / s1 98 / s2 87 episodes; "mix" = the three pooled. HF: Kavin60606/so100_bowl_train, so100_bowl_val_{s0,s1,s2,mix}.
• Policy: MolmoAct2 (Qwen3-4B + SigLIP2 + flow action expert), LeRobot 0.6.2, LoRA r64 on the VLM, full action-expert training, chunk 30 @ 30 Hz, absolute joint degrees, batch 16, 3000 steps, seed 1000.
• Checkpoint zoo (21): R1 100 % data every 250 steps to 3000 (12); R2/R3/R4 at 5/15/40 % stratified data at 1500 and 3000 (6); R5 frozen VLM at 1500 and 3000 (2); zero-shot lerobot/MolmoAct2-SO100_101-LeRobot (1).
• Critical intervals (commit → target committed; grasp → lifted; release → settled), three sources: sim-state GT on every episode; single-sample Gemini-2.5-pro on 50 episodes per split (the paper's protocol); median of three Gemini-2.5-pro samples on every episode.

2. Closed-loop protocol and what happened

The CI-MSE paper assumes deployment with temporal ensembling (query every step, execute the mean of the last H = 8 chunks) and builds its "+ens" metric to match. We ran that first: 0–10 % success on every checkpoint, videos show reach-then-hover. Averaging chunks sampled from a policy trained on multimodal demonstrations lands between strategies. Compute was also prohibitive: one 5B forward per env-step, ~70 min per 100 episodes on an A100.

Headline protocol was switched to MolmoAct2's native deployment: execute the full 30-step chunk open-loop, then re-query. 200 episodes per checkpoint (two seed blocks; the two blocks agree within ±5 pts). Temporal ensembling was kept as a control on 16 checkpoints. Offline dumps are independent of the controller.

3. Results

3.1 Closed-loop success (open-loop, 200 episodes)

run	checkpoints → SR
R1 100 %	250: .06 · 500: .12 · 750: .15 · 1000: .10 · 1250: .16 · 1500: .16 · 1750: .18 · 2000: .17 · 2250: .21 · 2500: .23 · 2750: .18 · 3000: .26
R2 5 %	1500: .16 · 3000: .28
R3 15 %	1500: .23 · 3000: .20
R4 40 %	1500: .20 · 3000: .16
R5 frozen VLM	1500: .14 · 3000: .12
zero-shot	.00

Temporal-ensembling control on the same seeds: 0.00–0.10 (R4@1500 highest at 0.10; R1@3000 0.04). First-failure mix at R1@3000: wrong cube 59 %, commit/grasp 17 %, carry/release 8 %.

3.2 Offline metric vs success

Spearman ρ (Kendall τ) against open-loop SR, n = 21. Columns are validation splits.

metric	s0 clean	s1 noisy	s2 varied	mix
raw MSE, all frames	−0.82 (−0.66)	−0.68 (−0.52)	−0.77 (−0.63)	−0.80 (−0.65)
CI-MSE + DTW1, GT windows	−0.71 (−0.55)	−0.47 (−0.35)	−0.79 (−0.64)	−0.73 (−0.57)
CI-MSE + DTW1, VLM single	−0.80 (−0.64)	−0.64 (−0.52)	−0.73 (−0.59)	−0.73 (−0.58)
CI-MSE + DTW1, VLM median-3	−0.78 (−0.62)	−0.72 (−0.56)	−0.76 (−0.60)	−0.77 (−0.62)
CI-MSE + ens8 + DTW1, GT	−0.63 (−0.47)	−0.20 (−0.15)	−0.65 (−0.50)	−0.42 (−0.31)
CI-MSE + ens8 + DTW1, VLM median-3	−0.64 (−0.46)	−0.31 (−0.21)	−0.67 (−0.49)	−0.55 (−0.41)

Bootstrap 95 % CI on ρ (episode resampling, mix): raw MSE [−0.86, −0.56]; ensembled GT [−0.55, −0.19]; ensembled VLM-3 [−0.67, −0.28]. Pearson on fine-tuned checkpoints only: raw −0.78, CI+DTW1 VLM-3 −0.81, ensembled GT −0.38.

Selection: picking the checkpoint with the lowest raw MSE or lowest CI+DTW1 (VLM) selects the true best (R2@3000, regret 0.00); the ensembled variants pick a 0.18 checkpoint (regret 0.10).

3.3 Controller mismatch

The only thing that separates the top and bottom rows of the figure is ensembling the predicted chunks before scoring. Ensembling is the paper's assumed deployment; on this policy it is destructive in the closed loop and uninformative offline. The non-ensembled metrics score exactly what open-loop execution runs, and they track it.

3.4 Interval source

GT, single-sample VLM and median-of-3 VLM give the same picture; median-of-3 is marginally best on the noisy split (−0.72 vs −0.47 GT). Agreement between Gemini and GT windows is low in IoU terms (grasp 0.2–0.3), yet the metrics agree, because window placement barely matters when windows cover ~95 % of frames (commit spans the whole approach).

3.5 Phase attribution

Spearman of per-window error against outcomes (mix, GT windows): commit-window raw error vs SR −0.88, grasp −0.61, release −0.75 — every window predicts overall success, none isolates its own phase. Wrong-cube rate correlates positively with lower error (+0.73) because better policies grasp more and therefore expose more colour mistakes. Phase-level diagnosis from offline error did not work here, as in v1.

3.6 Data scale and PEFT family

R2 5 % (0.28) ≥ R1 100 % (0.26) > R3 15 % (0.20) > R4 40 % (0.16) at 3000 steps. Spearman of metrics within this family: raw −0.40, ensembled +0.80 (wrong sign). The family is inside noise for SR and no metric resolves it. Frozen-VLM R5 (0.12–0.14) sits below LoRA R1 at the same steps and every metric places it correctly.

4. What this says about CI-MSE

• The critical-interval restriction is harmless but adds nothing when the task's "critical" phases span the episode. It should help most on long episodes with short decisive moments; this task is not that.
• DTW with W = 1 changes nothing measurable.
• The ensembling term is the risky part: it hard-codes a deployment controller into the metric. For a policy deployed open-loop it degrades ranking from ρ ≈ −0.8 to −0.4/−0.55.
• VLM annotation is good enough; no need for privileged state.
• Plain MSE remains the baseline to beat, and in two regimes (clean v1, multimodal v2) it was not beaten.

5. Caveats

• Success ceiling is low (0.28) and R1's middle checkpoints are within seed noise of each other, so the rank correlations are bounded by SR noise as much as by metric quality.
• Open-loop execution used the full 30-step chunk; other re-query horizons were not swept.
• Gemini annotation began on a Vertex express key that rate-limited and stalled; it was completed through OpenRouter with the same model (gemini-2.5-pro). Single-sample windows exist for 50 episodes per split.
• Zero-shot is included as the low anchor (SR 0, raw MSE 950 deg²); all correlations are also reported as Pearson on fine-tuned checkpoints only.
• Training loss plateaued at ~0.05 from step 500; the task (three cubes, colour grounding, DR, 2.8k demos) is hard for this data budget.

6. Cost and compute notes

8×A100-40GB for ~14 h (training 4.3 h, rollouts + dumps ~9 h). Temporal-ensembling evaluation is GPU-bound at ~70 min per 100 episodes; open-loop ~10 min; a full-frame dump ~35 min per split. An 8×H100 box was rented and dropped; an H100-specific cuDNN attention crash in the MolmoAct2 vision tower was worked around by removing the cuDNN SDPA backend.

7. Artefacts

• Kavin60606/cimse-so100-bowl-molmoact2-ckpts — 6 full checkpoints + 14 deltas (bit-exact reconstruction via ckpt_delta.py).
• Kavin60606/cimse-so100-experiment-v2 — results/ (per-episode closed-loop logs, videos), preds/ (63 prediction dumps), intervals/ (GT, VLM, VLM-3), code_v2/, report_v2/ (this report, summary.md, table.csv, correlation.json, scatter.png).
• Datasets: Kavin60606/so100_bowl_train, so100_bowl_val_{s0,s1,s2,mix}.
• Code: fd-studio/eval/sim_so100_v2/ (env, oracles, generation, training, eval, dump, correlation).