51²è¹Ý

The DARPA Learning Introspective Control (LINC) program is developing AI-driven technologies that allow robots to detect unexpected conditions, learn new dynamics, and adapt controllers to maintain safe operation while making progress towards mission goals—even under conditions of high uncertainty. In field tests with ground vehicles and cranes on moving platforms, our system, Adaptive Control with AI (ACAI), has consistently exceeded safety benchmarks, including surviving destabilizations four times stronger than anticipated. ACAI’s innovations include: a safety controller that guarantees stability against worst-case disturbances using ISAACS reachability methods, a functionality controller that preserves operator intent with Gaussian Process learning, and a repeated LQR approach that stabilizes crane payloads on moving platforms. Together, these methods provide a minimalistic, efficient, mathematically grounded system that shifts the burden of resilience from human operators to the robotic system itself. This talk presents the challenges we faced and lessons learned as we developed ACAI over the past three years.

Here is a zoom link for the same: https://zoom.us/j/98825797076?pwd=nj6WsQGzuaz6lGvySG7sUpiinLJyUM.1