GPU experiments 2026-10-02 — WorldGuard and SmolVLA

WorldGuard ensemble comparisons and SmolVLA fine-tuning on SO100 PickPlace recordings, with held-out metrics, pinned inputs and reproducible model artifacts. GPU measurements use an RTX 4090 D (24 GB).

• WorldGuard: 36 trained numerical, MuJoCo object and real SO100 state/visual ensemble members, with calibration, normalization/PCA, predictions, logs and exact source identities.
• SmolVLA: 5000 updates on the pinned real SO100 dataset; dev-selected checkpoint at step 3750. Held-out normalized action MAE decreased from 0.673166 to 0.245984 (63.46%) across five held-out episodes. This is action reconstruction, not hardware task success.
• Batch-one, four-thread image-in-memory to 50x6 action chunk inference: P50/P95 229.26/236.90 ms, excluding video decoding, transport and actuation.
• Camera-layout stress: the original profile produced 47/500 unsafe selections. Profile-matched dev/cal recalibration changed the same predictions to slower 1.6-second selections and observed 0/500 on that existing stress set; XY coverage remained 94.0%. This recovery is not a new independent deployment test.
• Negative comparisons, strong physical/state baselines, actual units and independent sample denominators are retained in the report.

Download artifact_index.json to verify the two bundle SHA-256 digests. Each bundle includes its own per-file ARTIFACT_INDEX.json, loading/model cards, attribution and licenses. The SmolVLA bundle is the selected trainable overlay and requires the pinned official frozen base/backbone; optimizer/RNG and full upstream weights are excluded. Original real videos remain at their dataset source.

Report: https://github.com/LancerLSY/sentinel-evc-lab/blob/main/docs/gpu_training_results.md
Model cards and loading: https://github.com/LancerLSY/sentinel-evc-lab/blob/main/docs/gpu_model_cards.md

Experimental profile. Live robot actuation, real object supervision and product permit/runtime integration remain separate work. Custom code/numerical-WorldGuard weights: MIT; upstream SmolVLA/SmolVLM2 and dataset: Apache-2.0, included with attribution.