EE - Trustworthy Physical AI: Learning Controllers You Can Prove, Not Just Tune
This project develops safe, trustworthy Physical AI by combining reinforcement learning with structure-preserving optimal control for robotic systems modeled as port-Hamiltonian (PH) systems, including robotic arms, mobile robots, and quadrotors. Instead of relying on opaque black-box policies, controllers are explicitly grounded in physics—using energy shaping and passivity to guarantee stability, while safety filters enforce hard constraints so the robot cannot violate actuator limits or enter unsafe regions during learning or deployment. Physics-informed neural networks capture the remaining unknown components, embedding known dynamics and stability conditions directly into training so the system learns efficiently without breaking physical laws. Students will build this trustworthy-AI pipeline end-to-end: modeling robot dynamics, coding nonlinear controllers and trajectory optimizers in Python, training safety-certified neural controllers, and validating them on real hardware against standard RL and linearized baselines. This offers a hands-on path to publishable research on Physical AI systems that are provably stable, verifiably safe, and optimal by design.
This lab is a fast-moving place to work at the intersection of Physical AI, nonlinear control, and robotics, where theory and hardware meet every week. The project on safe, trustworthy Physical AI is important because future robots must be not only intelligent, but provably stable and reliably safe when operating around people in the real world. In parallel with this project, the group is exploring physics-informed neural networks for complex dynamical systems, structure-preserving reinforcement learning for port-Hamiltonian robots, and safety-critical control with formal guarantees, creating a tightly connected set of efforts that feed into each other. Students join an environment where they can write code, run robots, and contribute to publishable research on how to make real-world AI systems we can actually trust.