Uncertainty-Driven Safe Continual Improvement of Generative Robot Policies

Generative robot policies, such as diffusion policies and vision-language-action models, have demonstrated impressive capabilities for solving complex, long-horizon tasks. However, deploying these models in our ever-changing real world, where robots must continuously adapt to novel situations without compromising safety, is very challenging. In my talk, I will argue for addressing these problems through a unifying lens: uncertainty awareness. I will first present how this enables achieving safety of diffusion policies without compromising task success. Then, I will discuss estimating uncertainty in end-to-end generative policies and leveraging it to 1) predict failures early during runtime and 2) acquire the most informative data for fine-tuning. Finally, I will conclude with ongoing work toward generalist policies that actively and continually learn from experts and their own experience.