Spring 2026 GRASP SFI: Stone Tao, University of California, San Diego, “Accelerating and Driving Robotics Progress with Simulation Integrated Robot Learning”

This was a hybrid event with in-person attendance in Levine 307 and virtual attendance…

ABSTRACT

Robotics research is advancing at an unprecedented pace, reaching scales that were difficult to imagine just a few years ago. However, as systems grow larger and more complex, effective research and engineering increasingly depend on having the right tools. In this talk, I introduce the concept of simulation-integrated robot learning and how this can serve both as a driver and an accelerator of progress in robotics.

As a driver, I argue how simulation can enable more efficient training, evaluation, and deployment of robotics models. I show how simulation can advance both large-scale robotics foundation models and task-specific policies for problems like cable routing. Beyond conventional use, I demonstrate how creative integration of simulation with robot learning, ranging from fast state resets for reinforcement learning to simulation-in-the-loop deployment, can improve performance and reduce sim-to-real dynamics gaps.

As an accelerator, principled simulation design and management can substantially reduce development time and bugs. I illustrate how the core ideas behind the ManiSkill3 simulation framework embody these principles, enabling rapid iteration and improved reliability. Examples range from making complex GPU-parallel simulation code look single-threaded to providing automatic sim-to-real sanity checks during development.