BAC: Block-wise Adaptive Caching Accelerates Diffusion Policy for Real-Time Robotics
Researchers propose Block-wise Adaptive Caching (BAC) to accelerate Diffusion Policy, a visuomotor model that is computationally expensive for real-time robotic control. Diffusion Policy suffers from high computational cost due to repetitive denoising steps, and existing acceleration techniques fail due to architectural and data divergences. BAC caches intermediate action features at the block level, adaptively updating and reusing them based on the observation that feature similarities have non-uniform temporal dynamics and block-specific patterns. An Adaptive Caching Scheduler identifies optimal update timesteps by maximizing global feature similarity, enabling lossless acceleration. The method addresses the redundancy across denoising steps while maintaining generation quality. The paper is available on arXiv under ID 2506.13456.
Key facts
- BAC is proposed to accelerate Diffusion Policy for real-time robotic control.
- Diffusion Policy has high computational cost due to repetitive denoising steps.
- Existing diffusion acceleration techniques fail to generalize to Diffusion Policy.
- BAC caches intermediate action features at the block level.
- Feature similarities exhibit non-uniform temporal dynamics and block-specific patterns.
- An Adaptive Caching Scheduler identifies optimal update timesteps.
- The scheduler maximizes global feature similarity.
- BAC achieves lossless action generation acceleration.
Entities
Institutions
- arXiv