Robot Learning: acting under safety constraints
DATE: 26 September
PLACE: Sala consiglio, 8th floor, Via Celoria 18
SPEAKER: Davide Tateo
Novel Deep reinforcement learning approaches can solve increasingly complex scenarios, such as Atari games, Chess, Go, and Starcraft. However, applying reinforcement learning to real-world environments is still extremely challenging, particularly when considering robotics tasks.
Differently from simulated scenarios and tabletop games, the real-world challenges prevent a naive application of Reinforcement Learning techniques for real robots.
Unfortunately, to achieve this goal, many practical problems need to be solved: safe and efficient exploration, trustability of the algorithm, and sample efficiency.
In this talk, we will discuss the crucial issues of Robot Learning, focusing in particular on how to act and plan under safety and task constraints.
We will present ATACOM, a Safe Reinforcement Learning framework that ensures the algorithm will sample only safe actions, by forcing the agent to act on the tangent space of the constraint manifold. We will show applications of this and other methods to both simulated and real-world robotics tasks.
Davide Tateo is a Postdoctoral Researcher at the Intelligent Autonomous Systems Laboratory, led by Prof. Jan Peters, Technical University of Darmstadt. He received his Ph.D. from Politecnico di Milano in 2019. His works are within many areas of Robotics and Reinforcement Learning, including Deep Reinforcement Learning, Planning, and Perception. His main research interest is Robot Learning, primarily focusing on high-speed motions, safety, and interoperability. He is co-author of the MushroomRL Reinforcement Learning library.
HOST: Matteo Luperto, Nicola Basilico