Safe RL (Nightmare Dreamer)
Nightmare Dreamer is a sample efficient, multi-agent approach to Safe RL. We learn a world model including safety constraints and use the model to learn both a safe policy and a control policy and in a hierarchical fashion use the model to look ahead to determine which policy to use. Our technique archives zero costs while maximizing rewards outperforming other model-free methods completely from image observation.
People
Dr Micah Corah
Interests: Aerial robotics, Active perception, and multi-robot systems with applications to aerial videography
