rogation/leboard-training-runs

LeBoard calibration: do dataset defects hurt a VLA?

LeBoard flags defects in LeRobot datasets: frozen video, a missing camera, actions that never change, desynchronised video, and more. This organisation holds a controlled experiment that asks the question behind every flag: does this defect actually lower the success of a policy trained on the data?

We take one clean dataset, break copies of it on purpose with one known defect each, fine-tune the same VLA on every copy with the same recipe, and compare success in simulation.

Setup

Base dataset	lerobot/libero @ a1aaacb, LIBERO-10 suite only: 379 episodes, 101,469 frames, 2 cameras, 10 fps
Variants	17 copies, each with one defect injected at a known rate: rogation/libero10-<variant>. Every copy carries calibration/injection_log.json listing the affected episodes and frames
Training copies	The same 17 re-encoded to H.264 with LeRobot's reencode_videos (crf 23): rogation/libero10-<variant>-h264. Decoding AV1 on the CPU made the GPU wait ~1 s per step; H.264 loads in ~0.006 s
Policy	lerobot/smolvla_base fine-tuned for 10,000 steps, batch 64, LeRobot 0.6.1, identical recipe for every run
Evaluation	lerobot-eval on LIBERO-10: 10 tasks × 50 episodes = 500 episodes per run, seed 1000, fixed initial states, so every run meets the same 500 starting scenes
Defect detection	Every variant scanned with LeBoard v2: rogation/leboard-episode-tables. All injected defects were found (recall and precision 1.00, or AUROC ≥ 0.94 for graded ones)

Results (2026-10-11, 10,000 training steps)

Clean reference: 42.2% success, mean of 5 seeds (43.8, 42.8, 40.6, 41.0, 42.6; spread ±1.3 points).

Variant	Defect, applied to 50% of episodes	Seeds	Success	Δ vs clean	Relative	Verdict
stuck-actions-r50	every action in the episode equals the first one	1	5.2%	−37.0	−88%	breaks the policy
desync-k3	video 3 frames behind the actions (all episodes)	3	38.3%	−3.8 ± 2.0	−9%	probably a small effect, not conclusive
action-noise-s30	Gaussian noise, 30% of each action dimension's std	1	40.8%	−1.4	−3%	no measurable effect
idle-start-r50	3 s of a still robot added before the episode	3	41.9%	−0.3 ± 1.5	−1%	no effect
freeze-r50	2 s of frozen video mid-episode	1	42.8%	+0.6	+1%	no measurable effect
wrist-absent-r50	wrist camera black for the whole episode	1	42.8%	+0.6	+1%	no measurable effect at 50%
blur-r50	two 2 s blurred windows	1	46.2%	+4.0	+9%	no measurable effect

How to read Δ. With one seed, a difference within about ±5 points is noise (spread between clean seeds plus the sampling error of 500 episodes). Desync and idle start were re-run with 3 seeds each; the ± is the standard error of the difference.

What this says so far

1. Stuck actions are severe. When half the episodes record an action that never changes, SmolVLA's success drops from 42% to 5%. LeBoard already flags this (static_actions).
2. An idle start does not hurt SmolVLA (3 seeds: −0.3 points). The literature's large idle effects are for single-step policies; SmolVLA predicts 50-action chunks.
3. Desynchronised video probably costs a little (3 seeds: −3.8 ± 2.0 points, about −9%), not yet conclusive.
4. Frozen video, blur, moderate action noise, and a wrist camera missing in half the episodes did not measurably hurt at these doses. The literature's large wrist-camera effect is for a camera missing from all training data; that dose is next.

Limits

• One seed per defect, one base dataset (LIBERO-10, simulation), one policy and recipe.
• Training is not converged at 10k steps: clean is 34.3% at 5k and 42.4% at 10k, and the SmolVLA paper reports 71% on LIBERO-Long with a much longer recipe on all four suites.
• Defects are synthetic and strong; real datasets have subtler ones.
• H.264 does not change the measurement: blur-r50 scored 46.2% from H.264 and 45.6% from the original AV1 videos.

Where everything is

• Variants: rogation/libero10-<variant> (AV1, as built) and rogation/libero10-<variant>-h264 (used for training). All are deliberately corrupted calibration data: do not train on them for any other purpose.
• Runs: rogation/leboard-training-runs, one folder per run with run.json (recipe, pinned data revision, success per checkpoint), training and evaluation logs, eval_info.json per task, evaluation videos and checkpoints.
• Episode tables: rogation/leboard-episode-tables.
• Harm model design: MODEL.md.
• Harm corpus (one row per run: LeBoard measures → relative drop in success): rogation/leboard-harm-corpus.

Defect reference: what hurts, how LeBoard measures it, and the evidence

The weight a LeBoard flag should carry comes from the literature first; our own calibration runs (SmolVLA, LIBERO-10, one seed, defect in 50% of episodes) are one extra data point, not a replacement. Status: ✅ implemented and validated on injected defects · 🟡 partial · ❌ not measured.

Defect	What the literature says hurts	References	How LeBoard v2 measures it	v2 status	Our calibration (Δ vs clean)
Static / no-op actions	Wrong action labels are the most damaging training defect: an all-zero first action in Bridge made OpenVLA predict zeros and freeze; filtering no-ops was "crucial"	OpenVLA	static_actions (every action dimension constant for the episode), static_action_dims (single frozen dimensions), zero_first_action	✅ recall 1.00, precision 1.00	−37.0 (stuck-actions-r50)
Idle starts and pauses	Idle actions make policies get stuck, worst for single-step policies; teams filter idle frames before training	Diffusion Policy, openpi DROID recipe, DemInf, FrameSkip	idle_start_s, idle_mid_s, idle_end_s, longest_idle_s (robot state moves < 1% of its typical range over a 1 s window); v1 idle_episodes	✅ idle start (recall 1.00); 🟡 mid-episode pauses measured but not yet validated	−0.3 ± 1.5 (idle-start-r50, 3 seeds: no effect)
Video / action desync	Timing errors cut success sharply at deployment: −30 points without latency matching; stale camera streams −65 to −77 points	UMI, LIBERO-VPro, Diffusion Policy	per_camera.sync.lag_frames: correlation of video motion with the best-tracking state dimension, pooled over episodes	✅ dataset lag exact (1 and 3 frames); 🟡 per-episode lag noisy	−3.8 ± 2.0 (desync-k3, 3 seeds: probably small, not conclusive)
Missing camera / dropouts	Losing a view, the wrist view above all, is among the largest effects: −30 points without the wrist camera in training; −64 to −75 at test time	robomimic study, MINT, LIBERO-VPro	camera absent (every frame near-blank), near_blank_frames (dropouts)	✅ absent camera (recall 1.00); 🟡 mid-episode dropouts not yet validated	+0.6 (wrist missing in 50% of episodes; literature: missing in all)
Operator quality / inconsistent actions	Demonstrations from less skilled or inconsistent operators lower success (robomimic Square BC 78.7% → 52.7%); action noise costs most on small datasets; smoothness and consistency predict success	robomimic, Belkhale et al., Sakr et al., RINSE, PSD	jerk_rms, sparc, path_length	🟡 detects random action noise (AUROC 0.95–1.00); ❌ inconsistency between demonstrators (different strategies for the same state) not measured	−1.4 (random noise only, action-noise-s30)
Frozen frames	A frozen frame paired with moving actions is a stale observation baked into training (same mechanism as test-time staleness above)	LIBERO-VPro (test-time staleness)	freeze_frames, longest_freeze_s (pixel-exact repeats ≥ 1 s)	✅ recall 1.00, precision 1.00	+0.6 (freeze-r50: 2 s in half the episodes)
Blur / darkness / lighting	Mostly a robustness issue: lighting changes cost ~15 points at test time, while varied images in training act as augmentation and help generalization	LIBERO-Plus, SEVO	blur_frames, dark_frames per camera (v1 rules)	✅ blur (AUROC 1.00); 🟡 darkness not yet validated	+4.0 (blur-r50)
Failed or bad-strategy demonstrations	Removing failed or brittle-strategy demos gives the largest curation gains (+15 to +35 points; CUPID: 44% → 84% on one task)	Demo-SCORE, CUPID, OpenVLA	— (length_ratio is only a weak proxy)	❌ needs success labels or policy rollouts	not tested
Low diversity / small size	Generalization follows the number of environments and objects (power law); varied lighting and backgrounds was "the single most important factor"; too much diversity can hurt pretrained VLAs	Lin et al., SEVO, Geometric Entropy	episodes, hours only	❌ scene / object / task diversity not measured	not tested
Task text and camera-role metadata	SmolVLA rewrote vague task strings and standardised camera names before pretraining on community data (51.7% → 78.3% with that data)	SmolVLA	—	❌ not measured	not tested
Constant state dimensions	A dimension that never moves gets ~0 std, so tiny deviations at deployment become huge normalised inputs	TheWisp/lerobot PR #233	static_state_dims in the summary	✅ reported	not tested
Metadata vs video duration	No published training-time measurement	—	v1 metadata/video duration check	✅ (not injected)	not tested

Reading the last column. Our calibration has one seed per defect: a gap within about ±5 points is noise. Where it differs from the literature (random action noise, frozen frames, desync, idle start), the setups differ too: SmolVLA predicts 50-action chunks (the literature's large idle effects are for single-step policies), the timing results in the literature are measured at test time rather than in training data, our action noise is random rather than an operator's structured mistakes, and 2 s of freeze in half the episodes is only ~7% of frames. Until more policies, datasets and doses are tested, the literature's ranking is the default.