closed-loop-checkpoints — policies + tactile tokenizer (T-Rex reproduction sim study)
Checkpoints from the closed-loop Isaac Sim study in
https://github.com/physercoe/trex-repro (docs/CLOSED_LOOP_EVAL.md).
Task: press-in peg-in-hole, RH56 hand, Isaac Sim 5.1 / IsaacLab 2.3.2,
tactile = ground-truth F6 wrenches. PyTorch .pt state dicts; architectures
in closed_loop/bc_policy.py / train_ppo.py / train_vqvae.py.
| file | role | eval SR (σ=20 mm clean / +disturbance) |
|---|---|---|
ppo3_champion_model_iter800_eval0973.pt |
blind PPO base ("champion", variant (b)); SR 0.973 on the clean task | 0.586 / 0.629 (frozen) |
ppo4_blind_model_final.pt |
blind fine-tune control (b′) | 0.785 / 0.766 |
ppo4_raw_model_0.pt |
raw-F6 tactile variant (c), ~iter 340 | 0.789 / 0.734 |
ppo4_tac2_model_final.pt |
VQ-token tactile variant (a) | 0.754 / 0.746 |
vqvae4_sim_tactile_vqvae.pt + vqvae4_norm_stats.npz |
tactile VQ-VAE tokenizer (30 Hz 16-step F6 windows → one-hot codes) + normalization stats | — |
All fine-tunes used identical PPO budget (400 iters, 512 envs) from the champion. Verdict of the study: tactile VQ-token features gave no robust advantage over the same-budget blind control in this setup — see the repo for the full table and caveats. License: MIT.