Robotics Seminar: Dr. Wanxin Jin, "Physics as the Backbone of Dexterity: Scalable Contact Simulation, Contact-Aware Control, and Physics-Grounded Learning."
Fri, Mar 13 · 2:00 PM - 3:00 PM
Abstract: Dexterous manipulation is fundamentally governed by physical contact. Rather than relying solely on larger datasets and end-to-end policies, I argue that contact-rich robot dexterity can be tackled more effectively by explicitly leveraging contact structure: where to interact, how contact evolves, and how motion unfolds under contact dynamics.
In this talk, I will present a physics-grounded view of robotic dexterity spanning contact simulation, control, world model learning, and human feedback. I will first introduce complementarity-free analytical contact modeling and simulation, which enable closed-form contact resolution, differentiability, and real-time (100 Hz) contact-implicit MPC for dexterous manipulation.
Building on this foundation, I will present ComFree-Sim, a GPU-parallel analytical contact physics engine that achieves linear runtime scaling with contact density and significantly improves throughput in dense contact-rich simulation and control.
I will then present a contact-interfaced hierarchical framework for geometry-aware long-horizon dexterous manipulation, which decomposes the problem into contact-intention learning and robust contact execution.
This structure enables data-efficient learning, robust performance, and zero-shot sim-to-real transfer. Next, I will discuss contact-aware world model learning, a differentiable vision-to-physics pipeline from sparse, contact-rich videos that unifies rendering and contact physics priors.
Finally, I will briefly describe our recent work on robust learning from human feedback, including robust reward alignment and direct fine-tuning of diffusion policies under corrupted feedback. Bio: Wanxin Jin is an Assistant Professor in the School for Engineering of Matter, Transport and Energy at Arizona State University.
His research lies at the intersection of robotics, control, and machine learning, with a focus on contact-rich dexterous manipulation, physics-grounded robot learning, and human-centered autonomy. Prior to joining ASU in 2023, he was a postdoctoral researcher at the GRASP Laboratory at the University of Pennsylvania.
He received his Ph.D. in Aeronautics and Astronautics from Purdue University in 2021. His work has appeared in leading robotics and machine learning venues, including T-RO, IJRR, RSS, ICRA, ICML, and NeurIPS.
He is an Associate Editor for IEEE Robotics and Automation Letters (RA-L). Meeting ID: 830 6832 2043Password: 746892