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 In particular, it relies on outdated observations to predict future actions, which can lead to inaccuracies.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 uncertainty 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.

Shifted Flow Policy

Action chunking predicts a fixed-length sequence of actions that all share the same timestep. Shifted flow assigns linearly increasing timesteps to each action within the prediction window to model future uncertainty.

Timestep \( t \) is continuously shifted as the action index \( k \) increases. By shifting \( t \) at the same rate as that at which it increases, a diagonal pattern emerges where noisy black actions are pushed towards white valid actions. Since the timesteps are simply shifted, the window can be initialized to have timesteps \(t_w = [1, (W−1)/W, \cdots , 2/W, 1/W]\). After generating the first action to execute on the environment from \( O_0 \), the window shifts to the right and adds noise to the end. This process is repeated to continuously generate more actions.

Results

Simulation Rollouts

Square

Threading

Coffee

Three Piece
Assembly

Hammer Cleanup

Mug Cleanup

Kitchen

Nut Assembly

Pick Place

Coffee Preparation

Comparisons with Action Chunking

Shifted Flow Policy outperforms existing approaches that rely on action chunking. This becomes noticable in contact-rich tasks where future actions become more uncertain. Shifted Flow Policy is able to account for this uncertainty.

Shifted Flow Policy is also highly efficient in the sampling process, only requiring one-step instead of multiple in the case of action chunking approaches.

Trajectory Visualization

Shifted Flow Policy treat future actions as more uncertain. This can be seen in the video above. Actions generated by Shifted Flow Policy become more random as the actions are further into the future.