T-Rex: Tactile-Reactive Dexterous Manipulation
Paper • 2606.17055 • Published • 1
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Artifacts from the independent T-Rex (arXiv:2606.17055) reproduction: github.com/physercoe/trex-repro.
Folders:
posttrain-ckpt-2000/ — weights (~8.5 GB) of our bounded 2000-step post-train
from the officially released midtrain checkpoint on a 144-episode subset of
the public zekaiwang/trex_dataset. Findings: docs/EVAL.md in the repo.closed-loop-checkpoints/ — policies + tactile tokenizer from the
closed-loop Isaac Sim study (press-in peg-in-hole, RH56 hand):
PPO champion (SR 0.973), tactile VQ-token variant (a), raw-F6 variant (c),
blind fine-tune control (b′), tactile VQ-VAE + norm stats.
Findings: docs/CLOSED_LOOP_EVAL.md in the repo.eval-data/ — the closed-loop eval JSON tree and f6 sample logs behind the
results tables.License: MIT (see the GitHub repo). The underlying T-Rex code/model are by the T-Rex authors (upstream: https://github.com/ZhuoyangLiu2005/T-Rex, MIT); the released midtrain checkpoint these derive from is theirs — cite arXiv:2606.17055.