NAT: Scaling Causal Neural Action Tokenizers Across Embodiments
Ahad Jawaid, Alexander Swerdlow, Yu Xiang
Preprint (2026)
TL;DR: A causal neural action tokenizer trained on 49k hours of robot, handheld-gripper, simulation and human action data: robots execute actions as their tokens are generated, and it outperforms FAST and prior neural tokenizers on unseen robot embodiments.




