Zero-Shot Reactive Obstacle Avoidance for Generative Robot Policies

2026-09-28

AI Summary

This study addresses the lack of real-time obstacle avoidance capabilities in generative robot policies during inference. To this end, it proposes NUDGE, a method that leverages signed distance field gradients to guide diffusion or flow matching policies for action generation. This approach enables training-free obstacle avoidance compatible with diverse action parameterizations without requiring retraining. By integrating differentiable geometry with a differentiable joint trajectory decoder, NUDGE achieves zero-shot, real-time reactive collision avoidance. Experimental results demonstrate that the proposed method preserves performance on the original task distribution while supporting multiple action formats and maintaining real-time responsiveness. Consequently, NUDGE provides an efficient solution for the safe deployment of generative policies in robotic manipulation tasks.

Abstract

We propose NUDGE (Nudge Update via Differentiable GEometry), a training-free obstacle-avoidance procedure that can be incorporated in any robot policy based on diffusion or flow matching, including diffusion policies and vision-language-action models. Our work injects gradients from a signed distance field, a function returning each point's distance to the nearest obstacle, into the policy at inference time to steer it away from obstacles. It supports any common action parameterization, from absolute or relative joint poses to end-effector poses, through a differentiable joint-trajectory decoder. Experiments show that NUDGE preserves the policy's task distribution and runs reactively in real time.

Problem

• obstacle avoidance
• generative robot policies
• diffusion policies
• zero-shot
• reactive planning

Innovation

• Zero-Shot Obstacle Avoidance
• Diffusion Policy
• Signed Distance Field
• Differentiable Geometry
• Training-Free