qqyang/zihiao_real_test

n0vtla_pretrain_scaling_study/ — mid-train (formerly Stage-1) human-ITW data-scaling study

Naming: the former "Stage-1 / pretrain" step is now called mid-train. Folder names, stage1_* metric names and the config name vtla_stage1_predictor_pretrain are unchanged so existing paths keep working. The study has three parts: mid-train (this folder), post-train and offline eval (n0vtla_scaling_posttrain/, below).

N0-VTLA Stage-1 predictor-grounding pretraining (paper Sec 4.2) on human in-the-wild (ITW) tactile data, trained on nested fractions {20, 40, 60, 80, 100}% of the corpus, then scored on one shared held-out episode set none of them trained on. Everything is warm-started from n0-vtla-base; only the tactile encoder projection, tactile predictor and a small recon head (~123M params) are trained (base policy frozen), 14000 steps, global batch 64, 8xH200.

Data split (why the comparison is fair)

• One seeded shuffle (seed 0) of the ITW raw corpus (39,426 episodes). 20/40/60/80% are prefixes of that same list, so splits are strictly nested (20 ⊂ 40 ⊂ 60 ⊂ 80).
• held_out = the complement of the 80% prefix: by construction in no training split. After a file-presence QC: 20%=7,885 · 40%=15,770 · 60%=23,656 · 80%=30,791 · held_out=7,731 episodes.
• 100% is a reference point from a separate, earlier run (n0-vtla_ts_pretrain/14000) trained on a different account's own pull of the corpus (mostly the same dates, not identical). Not drawn from the same shuffle; the held-out episodes were almost certainly unseen by it, but read the 100% point as "comparable distribution", not "same pool".
• Each run's own training loss is NOT a scaling metric: all fractions train the same 14000 steps, so a small pool is recycled more often and its training loss is biased low by memorization.

Checkpoint locations

Fraction	HF (Stage-1 trainable-only delta)	VISION delta (.../yqq/N0-VTLA_scaling_extract/checkpoints/vtla_stage1_predictor_pretrain/)	VISION merged base+delta (post-train ready, .../yqq/data/n0vtla/)
20%	n0vtla_pretrain_scaling_study/20pct/14000/	stage1_online_20pct/14000	n0-vtla-base_plus_stage1_scaling_20pct_14000
40%	n0vtla_pretrain_scaling_study/40pct/14000/	stage1_online_40pct/14000	n0-vtla-base_plus_stage1_scaling_40pct_14000
60%	n0vtla_pretrain_scaling_study/60pct/14000/	stage1_online_60pct/14000	n0-vtla-base_plus_stage1_scaling_60pct_14000
80%	n0vtla_pretrain_scaling_study/80pct/14000/	stage1_online_80pct/14000	n0-vtla-base_plus_stage1_scaling_80pct_14000
100% (ref)	n0-vtla_ts_pretrain/14000/	stage1_online_100pct_phailab/14000	n0-vtla-base_plus_stage1_human_14000

Delta = model.safetensors + optimizer.pt + metadata.pt (trainable params only). Post-training needs a complete model, so each delta is merged onto n0-vtla-base into one 8.25 GB model.safetensors (not on HF; VTLA_PRETRAINED_CHECKPOINT points at the merged dir).

held_out_eval_results/ — one sub-directory per evaluation protocol

Every sub-directory holds <label>.json (raw per-checkpoint report, label = 20pct/40pct/…), summary.csv (one row per checkpoint) and logs/eval_<slurm-job>.log (per-batch progress + final report). All protocols score the same held_out set with the same 200 shuffled batches (batch 64, ~11.8k valid frame samples; the loader yields one sample per FRAME, so a full pass is ~74k batches — infeasible — and a fixed random subsample is used). Lower is better for losses.

Directory	What it looks at
v1_batchnce_recon/	The Stage-1 training objective itself, evaluated on held-out. mean_stage1_nce: symmetric InfoNCE (temperature 1) between the predicted future-tactile latent z (mean-pooled, L2-normalised) and the real future latent z*, negatives = the other ~59 valid samples in the same batch. Chance = ln(59) ≈ 4.08. mean_stage1_recon: L1 between the recon head's 8x8 map (from z) and the 8x8-pooled future tactile field. mean_stage1_total = nce + 0.5·recon (the training loss).
v2_poolretr_reconbase/	Same forward pass, more sensitive readouts. retrieval: each predicted latent (query) is ranked against ALL ~11.8k real future latents in the pool (not just its batch) → top-1/5/10/100 accuracy, MRR, median rank (both i2t and t2i), pool_nce, mean_pos_cos vs mean_neg_cos (alignment margin), chance_top1=1/pool. recon_baselines: the recon head's L1 next to trivial predictors (all-zero, scalar mean, pool-mean field), overall and on contact cells (target above the 90th percentile), plus skill_vs_pool_mean_field = 1 − L1_model / L1_baseline (>0 means better than predicting the average map).
v3_task1ptdata_retr/	The same retrieval readouts on post-train (Task1 tuberack, robot wetlab) data, unseen by every Stage-1 checkpoint (they only trained on human ITW): a cross-embodiment transfer test in the post-train domain. Checkpoints are the merged base+Stage-1 models plus base = n0-vtla-base with no Stage-1 (reference). Splits: train (55 episodes, every 5th frame, capped at 6000 samples) and val (13 episodes, every 2nd frame), scored per split and pooled (train+val in one joint pool). summary.csv/json hold i2t/t2i top-1/5/10/100, MRR, pool_nce and mean positive cosine with episode-clustered bootstrap 95% CIs and paired differences vs base; embeddings/ has the raw normalised (z, z*) per checkpoint/split for re-pooling.

Results and interpretation

v1 (done). Held-out stage1_nce (lower is better): 20%=3.8001 · 40%=3.8021 · 60%=3.8050 · 100%(ref)=3.7995; stage1_recon ≈ 0.0034–0.0040 for all. Differences (≤0.005 nce) are non-monotonic and within noise: this protocol shows no data-scaling trend. Adding 80% (stage1_nce=3.8014) does not change this. Contributing factors: temperature 1 with cosine logits compresses InfoNCE's dynamic range (achievable floor ≈ 3.1 vs chance 4.08, and all models sit at ~3.80, i.e. close to chance), and the 8x8 recon L1 is tiny/saturated (contributes ~0.002 of the total). Training curves plateau at the same ~3.79–3.80. The v2 protocol was added to test whether a more sensitive readout separates them.

v2 (done, all five points). Pool retrieval on the 11,772-frame held-out pool (chance top-1 = 0.0085%): i2t top-1 = 1.42 / 1.35 / 1.17 / 1.26 / 1.25 % for 20 / 40 / 60 / 80 / 100%; i2t top-10 = 7.8 / 7.3 / 6.9 / 7.2 / 7.1 %; t2i top-1 = 3.54 / 3.19 / 3.20 / 2.78 / 2.57 %; pos−neg cosine margin ≈ 0.30 for all. The predictor clearly learns (top-1 ≈ 150x chance) but shows no data-scaling gain — 20% is nominally best; the gaps are within ~2 sigma with non-independent adjacent frames, so read it as a plateau, not as "more data hurts". Recon: the head's L1 (0.0034–0.0040) is ~7x worse than predicting all zeros (0.00052) overall and on contact cells (0.0047–0.0051 vs 0.0025); the target field is near zero (mean ≈ −1e-5), so v1's small recon value was not evidence of good reconstruction.

v3 (done). Task1 post-train-data retrieval (i2t; chance top-10 = 0.18% train / 0.33% val / 0.12% pooled; CIs = episode bootstrap, in summary.csv/json):

ckpt	train top-10	train MRR	val top-10	pooled top-10	pooled pool_nce
base (no Stage-1)	0.18%	0.0016	0.62%	0.14%	9.031
20%	2.87%	0.0083	2.93%	2.34%	8.890
40%	3.11%	0.0090	2.80%	2.63%	8.889
60%	3.18%	0.0108	3.13%	2.58%	8.884
80%	3.24%	0.0094	3.13%	2.69%	8.890
100% (ref)	3.31%	0.0094	4.17%	2.49%	8.883

1. Stage-1 clearly helps in the robot domain. base retrieval is at chance (train top-10 0.18% = chance; pool_nce 8.584 ≈ ln 5469 = 8.607). Any Stage-1 checkpoint lifts train top-10 ~16–18x and MRR 5–7x, and lowers pool_nce by 0.145; paired 95% CIs exclude 0 (val and pooled agree in direction).
2. Extra data adds little, and unstably. Train top-10 rises monotonically with data (+0.24 / +0.31 / +0.37 / +0.44 pp vs 20%; CI excludes 0 for 40/80/100%), but the absolute gain is ~0.4 pp, per-checkpoint CIs overlap heavily, val (13 episodes) shows no significant differences and pooled top-10 is non-monotonic. Read it as a weak trend on the train split, not evidence of gains beyond ~20%.
3. Absolute transfer is weak (top-1 0.03–0.4%, positive cosine ~0.15 vs ~0.32 in-domain human held-out).

Overall. On human held-out data no metric shows a data-scaling benefit; on robot data Stage-1 itself matters a lot but 20→100% adds only a weak, unstable difference. These are proxies; post-train and its offline evaluation follow below, and real-robot rollouts (hardware team) decide.

---

n0vtla_scaling_posttrain/ — post-train of the mid-train checkpoints + offline eval

Post-train runs

Each merged mid-train checkpoint (20/40/60/80%) was post-trained on two robot tasks, smoke_test_v2 (cap-to-tray, 59 train episodes, hand command is essentially two fixed poses: open / grasp) and Task1 (tube-rack hole transfer, 55 train episodes): 8 runs, global batch 64, 8xH200, 10,000 steps (config-default lr schedule, peak 2e-5, warmup 500, cosine decay over 20,000 steps, so the 10k checkpoint is at the same schedule state as the 10k checkpoint of the earlier 20k-step baseline/ours runs), ~3 h each. Real-robot rollouts are evaluated by the hardware team.

<exp>/<step>/ with <exp> = {sv2,task1}_posttrain_s1_{20,40,60,80}pct_10k and steps 5000 / 9999 / 10000, each with model.safetensors, metadata.pt, assets/ (no optimizer.pt; that stays on VISION).

Reference points (0% / 100%)

• Task1: n0vtla_wetlab_posttrain/post_train_task1_vision8gpu/{baseline,ours}/10000 (official base / Stage-1-human-14000 base, 20k-step runs, step 10000).
• smoke_test_v2 100%: n0vtla_wetlab_posttrain/human_data_post_train_v2/checkpoint_10000 (lab, 6 GPUs).
• smoke_test_v2 0%: n0vtla_wetlab_posttrain/checkpoint_20000 = official base post-trained on the OLDER smoke_test data (asset wetlab_v2_train, own norm stats). Not a like-for-like reference.

offline_eval/ — one sub-directory per protocol (each: per-checkpoint <task>_<pct>.json, summary.csv, logs/)

Only the final step-10000 checkpoint is evaluated; smoke_test_v2 checkpoints on canonical_wetlab_smoketestv2_dev (7 episodes), Task1 checkpoints on canonical_wetlab_task1_val (13 episodes), each with its own training normalization. Cells: {0,20,40,60,80,100}% x {smoke_test_v2, Task1}.

Directory	What it looks at
heldout_action_loss/	The flow-matching action loss the policy is trained with (denoising-field MSE in normalized space), computed with no gradient on held-out episodes: offline imitation fit, not closed-loop success. Every 5th frame, mean of 4 noise/time draws per frame with row-seeded draws shared across checkpoints. mean_action_loss_micro/macro + episode_se; by_group splits it into xyz / rot6d / hand (6 Revo2 motor commands); by_horizon gives the loss per step of the 50-step chunk.
open_loop_action_error/	Samples a 50-step chunk (10 denoising steps) per scored frame and compares it with the demonstration in physical units: xyz L2 error in mm, hand command MAE in raw counts (0-1000), rot6d MAE. Every 10th frame, 2 samples per frame. Baselines: predict zero motion (xyz/rot6d) and the training-set mean hand command. `metrics.<sample

Results and interpretation

Flow loss (total, mean ± episode SE): smoke_test_v2 20/40/60/80% = 0.0141 / 0.0137 / 0.0139 / 0.0138 (±0.0021); Task1 0/20/40/60/80/100% = 0.0556 / 0.0548 / 0.0554 / 0.0557 / 0.0557 / 0.0556 (±0.0055-0.0060). Sampled xyz error (mm): smoke_test_v2 20/40/60/80% = 11.63 / 11.62 / 11.54 / 11.40 (±0.4; zero-motion baseline 36.25); Task1 0/20/40/60/80/100% = 15.76 / 15.28 / 14.99 / 15.18 / 15.05 / 15.39 (±1.6-1.9; zero-motion 21.22). Hand MAE is flat at 13.1-13.3 (smoke_test_v2) and 16.4-17.2 (Task1) counts vs 228 / 280 for the mean-command baseline.

1. No measurable effect of mid-train data quantity (20-80%) on either offline metric: differences are within one standard error; paired per-episode differences vs 20% are at most ~3% of the loss and not monotonic.
2. Not even mid-train vs none is distinguishable on Task1 (0% vs 100%: loss 0.0556 vs 0.0556, xyz 15.76 vs 15.39 mm), although the same checkpoints differ hugely in tactile-latent retrieval (v3 above): post-training washes the difference out offline.
3. The policies are genuinely useful: sampled xyz error is ~30% (Task1) / ~68% (smoke_test_v2) below the no-motion baseline.
4. smoke_test_v2 0% and 100% are not comparable in total loss / rot6d (loss 0.849 / 0.430 and rot6d open-loop error 0.335 vs 0.0005-0.0024 for 20-80%; the two rot6d errors are almost identical, which points to a rotation-representation mismatch between those earlier checkpoints and the current pipeline - not verified). Their xyz / hand terms are fine for the 100% (11.71 mm, 13.0 counts). The 0% (older data) fails on smoke_test_v2 (38.8 mm, worse than no motion).

Limits: 7 / 13 held-out episodes, one seed per fraction, final checkpoint only, and offline fit is not closed-loop success.