Self-Guided Action Diffusion

2025-08-16

AI Summary

Existing diffusion-based robotic policies enhance action diversity during inference via multi-sample search, but incur prohibitive computational overhead that scales quadratically with sample count. This work proposes Self-Guided Action Diffusion (SGAD), a novel inference framework that replaces the conventional bidirectional decoding scheme with a dynamic proposal distribution—conditioned on prior decision information—at each denoising step. By leveraging task-relevant priors to guide stochastic action generation, SGAD preserves policy consistency and real-time responsiveness while drastically reducing inference latency. Experiments on high-dynamics simulated locomotion and manipulation tasks demonstrate that, under constrained sampling budgets, SGAD achieves up to 70% higher success rates compared to baseline diffusion policies and approaches the performance of oracle-optimal controllers. These results establish SGAD as an efficient, robust paradigm for generative robot control—bridging the gap between expressive diffusion modeling and deployable real-time execution.

Technology Category

• Search and Optimization: Sampling/Simulation-based Search
• Computer Vision: Diffusion Models for Vision
• Intelligent Robots: State Estimation

Application Category

• Search and Retrieval-Augmented AI: Agentic search
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Recent works have shown the promise of inference-time search over action samples for improving generative robot policies. In particular, optimizing cross-chunk coherence via bidirectional decoding has proven effective in boosting the consistency and reactivity of diffusion policies. However, this approach remains computationally expensive as the diversity of sampled actions grows. In this paper, we introduce self-guided action diffusion, a more efficient variant of bidirectional decoding tailored for diffusion-based policies. At the core of our method is to guide the proposal distribution at each diffusion step based on the prior decision. Experiments in simulation tasks show that the proposed self-guidance enables near-optimal performance at negligible inference cost. Notably, under a tight sampling budget, our method achieves up to 70% higher success rates than existing counterparts on challenging dynamic tasks. See project website at https://rhea-mal.github.io/selfgad.github.io.

Problem

• Improving generative robot policies via efficient action sampling
• Reducing computational cost in bidirectional decoding for diffusion policies
• Enhancing success rates in dynamic tasks with tight sampling budgets

Innovation

• Bidirectional decoding for cross-chunk coherence
• Self-guided action diffusion for efficiency
• Prior decision-guided proposal distribution