ReAT: Residual Action Tokenization for Autoregressive Robot Policies

This research proposes ReAT, a residual action tokenization method that enables each token to explicitly refine the current action reconstruction. By using decoded reconstruction as feedback, ReAT closes the loop between encoding and decoding, allowing tokens to respond to remaining action errors. Evaluated across 19 manipulation tasks, ReAT policies outperform baselines in task success, with even 4-token ReAT surpassing 8-token baselines on LIBERO at half the token budget.