Dynamic Neural Koopman Distillation for Fast Robot Control Using Diffusion Models

This research proposes a Dynamic Neural Koopman (DNK) distillation framework to accelerate robot control using diffusion models. By distilling multistep diffusion inference into a single forward pass, the method reduces inference latency while maintaining competitive performance on robot control benchmarks. The approach is evaluated on tasks involving locomotion, state-based, and image-based manipulation, showing improved efficiency compared to one-step policy baselines.