What we work on

Unmanned systems need enough onboard intelligence to cut human workload rather than add to it under SWaP constraints. We work on reinforcement learning, human-guided learning, multi-objective and fair policy optimization, and visual-language-model agents that are deployable in physical robotic systems. Recent work covers reward learning from human ratings, learning from failure, interpretable policy distillation, visual-language-action models, and inference-time policy alignment at deployment.

The lab has a long record in distributed coordination, consensus, containment, and event-triggered control. Current work is in safety-critical control under input constraints, safety certificate, and energy-efficient flexible leader–follower formation.

Connecting physical systems through the cyber domain creates capability and exposure at the same time. We study where the vulnerabilities are, what they cost operationally, and how to predict and mitigate them. Related work covers fault-tolerant control of uncertain multi-agent systems and resilient learning under corrupted data.

Research moves into practice only if it is validated on real hardware in real conditions. The lab designs its own small aircraft, including the airframe, the flight control board with onboard computing, the software, and the sensing modules. We build the benchmarks, datasets, and test protocols that make a capability claim checkable by someone outside the team.

The lab builds its own small aircraft end to end: the frame, the flight control board with onboard computing, the software stack, and the sensing modules. One team controls the compute carrier and the software, which is what makes a claim about onboard timing and autonomy measurable rather than estimated.

We gratefully acknowledge support from the Office of Naval Research, the Army Research Office, the Air Force Research Laboratory, AFOSR, CPS Energy, and NSF.