Free-View Robot Manipulation: Visuomotor Policy by Calibration Diffusion

ICLR 2026 Conference Submission14608 Authors

Keywords: Diffusion Policy; Free View;Visuomotor Robot Manipulation;

TL;DR: The robot complete the manipulation under Free-View

Abstract:

Visuomotor policies have demonstrated great potential in robot manipulation tasks.
However, current robot manipulation tasks are often observed from fixed viewpoints.
Once the viewpoints change, the trained policy becomes ineffective.
This limitation curbs the generalization of robot manipulation and impedes its application.
To address this issue, we make a comprehensive study by presenting novel free-view manipulation tasks that enables the robot to perform actions from any viewpoint.
Firstly, we construct a free-view dataset, which encompasses 8 tasks with over 5,000 episodes sourced from the ISAAC SIM simulation environment.
Each episode records robot manipulation behaviors from different viewpoints.
Secondly, we propose a calibration diffusion policy, which utilizes an additional calibration network to enhance the adaptability of the diffusion policy to different viewpoints.
In particular, we adopt two-stage curriculum training to make the calibration diffusion policy converge rapidly.
Finally, we conduct a wealth of experiments on the free-view dataset.
The obtained results demonstrate the effectiveness of the calibration diffusion policy.
This also means that we have built a new benchmark for free-view manipulation.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 14608