Spatial Attention: Adapting Execution Horizons for Diffusion Policies via Observation Sensitivity

2026-07-06

AI Summary

Diffusion policies with fixed execution horizons struggle to balance responsiveness and computational efficiency in dynamic environments. This work proposes a dynamic horizon scheduling method based on spatial attention, which quantifies state sensitivity via the squared Frobenius norm of the expected observation gradient and adaptively adjusts the execution horizon of action chunks in real time—shortening it under high sensitivity to enhance responsiveness and extending it under low sensitivity to reduce computational cost. The approach integrates seamlessly into the diffusion policy framework while improving robustness to perturbations. Experimental results demonstrate that, without altering the average execution horizon, the proposed method significantly outperforms fixed-horizon baselines in task success rate, achieving a synergistic optimization of responsiveness and computational efficiency.

Abstract

Sampling action chunks via generative models has become a widely adopted methodology for robotic learning from demonstration. However, existing methods often struggle to balance responsiveness and computational cost because they execute each action chunk for a fixed execution horizon. In this paper, we adaptively adjust the execution horizon of sampled action chunks, balancing responsiveness and computational efficiency. We introduce Spatial Attention -- defined as the expected squared norm of the gradient of the action log-likelihood with respect to the observation -- which indicates the sensitivity of the policy's action distribution to variations in the observation. We show that, under a fixed budget of chunk samplings, the execution horizon that minimizes the cumulative likelihood drop induced by disturbances decreases as Spatial Attention increases. By forecasting future Spatial Attention values alongside the action chunk, our framework dynamically assigns shorter execution horizons to phases with high Spatial Attention, and longer horizons to phases with low Spatial Attention. Experiments on standard and perturbed tasks, in both simulation and on a real robot, show that our method significantly improves success rates over fixed-horizon baselines while maintaining the average execution horizon.

Problem

• diffusion policies
• execution horizon
• robotic learning
• responsiveness
• computational cost

Innovation

• Spatial Attention
• Diffusion Policies
• Adaptive Execution Horizon
• Observation Sensitivity
• Robotic Learning