Research Analyzer
← Back ICRA 2026

Shifted Flow Policy: Uncertainty-Aware Time Reparameterization for Visuomotor Learning

Dasom Ahn, Chanhyuk Jung, Joonki Baek, Sungkeun Yoo, Byoung Chul Ko

PDF

AI summary

Key figure (auto-extracted from paper)
Shifted Flow Policy replaces action chunking with time reparameterization to model increasing uncertainty, delivering higher success rates and faster inference for robotic manipulation.
Shifted Flow Policy Visuomotor Learning Action Chunking Flow Matching Robotic Manipulation Time Reparameterization

Problem

Action chunking reduces compounding errors in visuomotor policies but relies on outdated observations to predict future actions, leading to local inaccuracies and ignoring real-time feedback.

Approach

The method linearly shifts time steps across a prediction window so each generated action is conditioned on fresh observations, explicitly modeling how uncertainty grows over time.

Key results

  • Outperforms state-of-the-art action chunking baselines on MimicGen and Push-T benchmarks
  • Achieves significantly faster inference latency (5.22 ms) and higher throughput
  • Demonstrates superior robustness in contact-rich and long-horizon manipulation tasks
  • Introduces a novel shifted flow formulation requiring only minor timestep encoder modifications

Why it matters

It offers robotics researchers and developers a more robust and efficient alternative to action chunking for building real-time visuomotor policies.

Abstract

Imitation learning for robotics often uses action chunking to mitigate the compounding errors associated with autoregressive policies. By predicting multiple future actions simultaneously, action chunking limits the accumulation of errors but introduces new difficulties. In particular, it relies on outdated observations to predict future actions, which can lead to inaccuracies. In this study, we propose Shifted Flow Policy (SFP), a simple yet effective alternative to action chunking. The SFP reparameterizes time by linearly shifting the time steps for future actions, thereby capturing the natural increase in uncer- tainty over time. This formulation allows each predicted action to be conditioned on up-to-date observations. Experimental results on the Push-T and MimicGen benchmarks demonstrate that SFP outperforms state-of-the-art action chunking methods across a variety of manipulation tasks by achieving higher success rates and faster inference. These findings suggest that shifted flow provides a robust and practical alternative to action chunking in visuomotor policy learning. Our code is available at https://shifted-flow-policy.github.io

Index terms

Imitation Learning Deep Learning Methods Motion and Path Planning

Related papers