RI Seminar: Andrea Thomaz : Robots Learning from Human Teachers

In this talk

1. 1:31

   Robots as teammates

   Thomaz wants robots to work beside people as teammates, responding appropriately in shared spaces instead of waiting for commands. She frames the challenge around ad hoc teams, where people coordinate, divide tasks, and learn on the job.

2. 6:02

   Learning goals from demonstrations

   Thomaz asks what a robot should learn from a human demonstration. In a box-opening example, Curie learned both the demonstrated movements and a model of what success looked like. That goal model let it recognize failure and practice adjusted movements, even when the original demonstrations failed.

3. 13:35

   Making human-guided learning more efficient

   Thomaz describes three approaches: adapting exploration to whether a teacher is paying attention, tuning a simulator from a single real-world example, and transferring tasks between tools using human corrections. The tool-transfer work reached over 83% accuracy within a task and about 30% across new tasks, rising to 50% when tasks were grouped by whether translation or orientation mattered more.

4. 39:50

   Applying robots in hospitals

   Thomaz co-founded Diligent Robotics to build Moxie, a mobile manipulator that supports clinicians with fetching, delivery, and room setup. In pilots across five Texas hospitals, Moxie can respond to electronic health record triggers, such as preparing a room after it is marked clean. The team tracks whether tasks reduce nurses’ workload and how quickly the robot can handle on-demand requests.