Siying Wang

Evolving Cache Schedules for Fast Diffusion Policy Inference

Siying Wang, Kangye Ji, Di Wang, Fei Cheng

Diffusion policies achieve strong visuomotor control by iteratively denoising action chunks, but repeated denoising makes real-time deployment computationally demanding. Cache-based methods reduce inference cost by reusing intermediate activations, but existing training-free sche…Read more