Georgia Chalvatzaki, a research group leader at TU Darmstadt and postdoc in Jan Peters’s Intelligent Autonomous Systems group, presents work on speeding up robot learning from demonstrations and learned models at the University of Toronto’s AI in Robotics Seminar.
Chalvatzaki describes the challenge of applying deep reinforcement learning to physical robots: exploration can be unsafe, and training often requires too many real-world interactions. Her group explores demonstrations, learned dynamics, and off-policy methods to make policy learning more practical.
The LAMPO approach learns a probabilistic representation of demonstrated movements and task context, then optimizes in a smaller latent space instead of directly searching over many movement parameters. It reuses experience from earlier policies with importance sampling and regularizes changes to the context distribution and policy. In robot demonstrations, reaching improved with additional interaction, and a pouring task learned to control the amount of sugar after starting with 100 demonstrations and 225 learning interactions.
MoPAC combines real transitions collected by an actor-critic method with simulated rollouts from an ensemble dynamics model. Model predictive control and value-function look-ahead guide the rollouts, whose transitions are added to the replay buffer for policy updates. In tests on simulated robot tasks and a physical OpenHand, Chalvatzaki reports faster learning than MBPO and SAC on valve rotation and finger gaiting.
In the discussion, Chalvatzaki says movement primitives are flexible but will not produce an optimal policy or generalize across every task. She recommends a library of simpler skills, combined by a higher-level controller, as a practical direction for complex and nonlinear tasks. She also notes that LAMPO needs demonstrations to initialize its latent representation.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Georgia Chalvatzaki, independent research group leader at TU Darmstadt, discusses robot learning and mobile manipulation research with PAL Robotics’ host, including work with TIAGo and TIAGo++.
Georgia Chalvatzaki, independent research group leader at TU Darmstadt, discusses her path into robotics, her research plans, and the role of TIAGo++ in her lab’s work.
