<ul><li>Diffusion policies are widely used in decision-making scenarios like robotics, gaming, and autonomous driving for learning diverse skills.</li><li>Existing diffusion policies could be sub-optimal due to limited demonstration data, leading to sub-optimal trajectories and failures.</li><li>A new framework called NCDPO has been introduced to address challenges in fine-tuning diffusion policies, enabling tractable likelihood evaluation and gradient backpropagation through all diffusion timesteps.</li><li>Experiments demonstrate that NCDPO achieves comparable sample efficiency to traditional methods and outperforms them in both efficiency and final performance across various benchmarks.</li></ul>

Fine-tuning Diffusion Policies with Backpropagation Through Diffusion Timesteps

Discover more