Abstract:
The discussed topic concerns reinforcement learning (RL) in robotics, where I try to address the issue of safety and interpretability. Today’s RL research is done mostly in simulated environments, but eventual deployment in real-world scenarios will require addressing the issue of assuring interpretability & safety of the trained RL agents. I have been working in collaboration on a novel approach termed state-planning policy RL. I will present some preliminary results obtained so far, and discuss future research prospects related to visual planning RL and offline RL.

свернуть