Resilient Autonomous Systems

Current autonomous systems can perform impressive tasks, from retrieving target objects in clutter to flying at high speed through a forest. However, no existing autonomous system can perform a high-level task over an extended period of time in an unstructured and changing environment, let alone provide guarantees for such behavior. Over time, prior assumptions encoded in training data and model structure are violated, inevitably resulting in system failure.

Hadas Kress-Gazit, Sibley School of Mechanical and Aerospace Engineering, is leading a team that includes researchers at Cornell, Brown University, and the Massachusetts Institute of Technology to develop a new science of resilient composition. This science is based upon inferring, verifying, composing, and repairing learned and model-based representations, which are designed to accurately predict the dynamics of the world and the effects of robot actions. In contrast to prior work on composing fixed models, the team is focusing on the science required to build and verify provable models directly from data and to repair those models online in response to dynamic environments. This approach draws inspiration from and leverages advances in program verification and synthesis, machine learning, estimation, robotics, control, optimization, perception, and graphics.

Creating autonomous systems capable of long-term self-reflection and self-repair will revolutionize the type of missions these systems can carry out successfully and safely. It will enable long-term autonomous missions where systems are deployed for months, not hours, in changing environments and conditions. 

Cornell Researchers

Funding Received

$7.5 Million spanning 5 years

Sponsored by

Other Research Sponsored by United States Department of Defense, Office of Naval Research