Discovery Machines: From Pixels to Physics
Talk Summary or Abstract: What if a machine could watch a video and write down the equations of motion? I'll discuss our recent work toward this goal - systems that learn to discover the hidden state variables and governing dynamics directly from raw visual data. These discovered Neural State Variables and Neural State Vector Fields not only reproduce trajectories but also reveal the structure of these dynamical systems: equilibrium, stability, and chaos. The results suggest a path toward a self-driven AI scientist.
Speaker Short Biography: Dr. Boyuan Chen is the Dickinson Family Assistant Professor at Duke University, where he directs the General Robotics Lab. He is affiliated with the Departments of Mechanical Engineering and Materials Science, Electrical and Computer Engineering, and Computer Science. He also serves as the Strategic Advisor for Robotics and Autonomy in the Dean's Office at Duke's Pratt School of Engineering.
Dr. Chen received his Ph.D. in Computer Science from Columbia University in 2022. Dr. Chen received his Ph.D. in Computer Science from Columbia University in 2022. His research focuses on building Discovery Machines - machines that learn, act, and collaborate by discovering how the world works. His lab develops the arc of Embodied Intelligence in which machines sense through multiple modalities, adapt by discovering and designing their own bodies and capabilities, and connect by understanding how humans and other machines think and act. By taking a full-stack approach that spans both the "body" and "brain" of intelligent systems, his group advances robotics, artificial intelligence, human-AI teaming, and AI for scientific discovery. Inspired by natural intelligence in humans and animals, his work explores new frontiers in adaptive, multimodal, and interactive autonomy.
Dr. Chen was named to ASME's Watch List in 2025. His work has received numerous media reports and has been featured in outlets such as the New York Times, Forbes, Fox, Fortune, Science, and the National Science Foundation. His research has been published in top-tier venues including Science Robotics, Nature Computational Science, NeurIPS, ICRA, and CoRL.
