Welcome to the Reliable Autonomous Systems Lab at MIT (REALM)! We design, analyze, and verify safe control systems. Our work lies at the intersection of control theory, machine learning, and formal methods, with a particular focus on safety in systems with nonlinear, high-dimensional, and difficult-to-model dynamics.
Some highlights from recent projects...
1 / 5
To enable safe control of up to 1024 distinct agents, each with nonlinear dynamics, we learn a barrier-function-based neural controller to ensure scalable collision-free navigation. This image shows 1024 nonlinear quadrotors executing a maneuver using this controller.
2 / 5
To enable safe control of up to 1024 distinct agents, each with nonlinear dynamics, we learn a barrier-function-based neural controller to ensure scalable collision-free navigation. This image shows 64 nonlinear cars executing a maneuver using this controller.
3 / 5
To solve reach-avoid control problems for complex systems with nonlinear dynamics, we develop robust Lyapunov/Barrier functions that certify the safety of a learned controller. This image shows a certificate for collision avoidance, which learns the outlines of obstacles as super-level sets
4 / 5
To safely control large-scale multi-agent systems, we propose to jointly learn the decentralized controllers with control barrier functions. The animation shows 1024 quadrotors controlled to deliver packages in a city, and avoid collision with each other.
5 / 5
We propose Density Constrained Reinforcement Learning, where the constraints are imposed on state density rather than value functions. State density has a clear physical intepretation and can express a wide variety of constraints. The image shows the electrical vehicle routing task in Manhattan, where the vehicle density at charging stations are constrained.