Evolving Cache Schedules for Fast Diffusion Policy Inference

TL;DR - EVO uses evolutionary search to optimize cache-refresh schedules for pretrained diffusion policies, accelerating visuomotor action generation without retraining. It achieves up to an 8.05× speedup while retaining near-full task performance.

• Searches globally across transformer blocks and denoising timesteps to identify reusable activations.
• Uses redundancy-aware initialization and target-conditioned early stopping to make offline schedule optimization practical.
• Plugs optimized schedules directly into existing diffusion policies.
• Reduces inference compute from 15.77G FLOPs to as little as 1.96G on manipulation benchmarks.