Fast Policy: Accelerating Visuomotor Policies without Re-training

Abstract

Diffusion models are increasingly employed in visuomotor policies to achieve promising performance of behavior cloning. However, the slow inference caused by iterative denoising is a notorious disadvantage, which greatly limits its application in resource-limited and real-time interactive robot systems. The prevailing strategy to this problem is distillation, but it still requires considerable resources to retrain a student model. To this end, we take another training-free view to develop a novel Fast Policy (termed FP), which can be regarded as a powerful and accelerated alternative to Diffusion Policy for learning visuomotor robot control. Specifically, our comprehensive study of UNet encoder shows that its features change little during inference, prompting us to reuse encoder features in non-critical denoising steps. In addition, we design strategies based on Fourier energy to screen critical and non-critical steps dynamically according to different tasks. Importantly, to mitigate performance degradation caused by the repeated use of non-critical steps, we further introduce a noise correction strategy. Our FP is evaluated on multiple simulation benchmarks and the comparison results with existing speed-up methods demonstrate our effectiveness and superiority with state-of-the-art success rates in visuomotor inference speed. The code is available at https://github.com/xwccchong/Fast-Policy