HiFlow: Tokenization-Free Scale-Wise Autoregressive Policy Learning via Flow Matching

We propose HiFlow, a tokenization-free coarse-to-fine autoregressive policy that operates directly on raw continuous actions via flow matching, eliminating the need for discrete action tokenizers. HiFlow constructs multi-scale continuous action targets from each action chunk via simple temporal pooling and is trained end-to-end in a single stage. Experiments on MimicGen, RoboTwin 2.0, and real-world environments demonstrate that HiFlow consistently outperforms existing methods including diffusion-based and tokenization-based autoregressive policies.

Citation: D. Yashima, K. Seno, S. Kurita, Y. Oda, and K. Sugiura, "HiFlow: Tokenization-Free Scale-Wise Autoregressive Policy Learning via Flow Matching", arXiv, 2026.
Download Paper