wy891/visuotactile-paper-artifacts

Visuotactile Paper Artifacts

Small trained artifacts and feature caches for this paper's visuotactile training pipeline.

Source

These artifacts come from the training pipeline in https://github.com/howard-lynn-ye/vr-teleop-isaacsim. Paper draft: https://github.com/howard-lynn-ye/vr-teleop-isaacsim/blob/main/paper/main.tex.

Pinned source commit for these uploaded artifacts: 8084ec32d37f07d57bd71e424d6144aae9e4f8b9.

Contents

• phase1/seed_42/best.pt, phase1/seed_42/last.pt: Phase 1 alignment encoder checkpoints.
• phase1/features_cache/: canonical Phase 1 DINOv2 feature caches and split manifest.
• phase3_3/seed_42/best.pt, phase3_3/seed_42/last.pt, phase3_3/seed_42/train.log: Phase 3.3 sim alignment artifacts.

Reproducibility

• Canonical Phase 1 data split: DATA_SEED=42.
• Phase 1 training script: https://github.com/howard-lynn-ye/vr-teleop-isaacsim/blob/8084ec32d37f07d57bd71e424d6144aae9e4f8b9/path_d/code/train/phase1_5_5seed.py.
• Phase 3.3 training script: https://github.com/howard-lynn-ye/vr-teleop-isaacsim/blob/8084ec32d37f07d57bd71e424d6144aae9e4f8b9/path_d/code/train/phase3_3_alignment.py.
• Phase 1 feature cache manifests record cache_version=2, n_rows=4583, top_n=30, and the shared sample_ids_sha256.

Critical Caveats

• Phase 1 checkpoint: best_ep=67, val_loss=2.5115, top5=0.6876. This post-restore retrain matches the original within epsilon.
• Phase 3.3 uses the held-out-scale split required by gamma_gates compliance: train on s15+s20 and validate on s25.
• These are paper artifacts, not a turnkey production model package. Preserve the split and script commit when comparing numbers.

Related Repositories

• SSVTP source corpus: https://huggingface.co/datasets/mlfu7/Touch-Vision-Language-Dataset
• Imputed force sidecars: https://huggingface.co/datasets/wy891/ssvtp-claude-imputed-force
• Unified v2t generator: https://huggingface.co/wy891/v2t-unified-ssvtp-diffusion