Beyond Reconstruction: What Matters in Action Tokenization for Robot Policies?

This research introduces ProAct, a novel action tokenizer training method that improves downstream predictability and robustness by augmenting reconstruction objectives. Evaluated across multiple benchmarks and real-world tasks, ProAct achieves average gains of 11.3 percentage points in rollout success and 21.8-36.7 percentage points in vision-language-action policies and robotic manipulation.