Context Aware Policy Adaptation: Towards Robust Safe Reinforcement Learning
Keywords: safe reinforcement learning, transfer learning
Abstract: Real world tasks are often safety-critical, requiring policies that respect safety constraints while also being able to safely adapt to novel situations. Typical safe reinforcement learning methods focus on adapting to shifts in the transition function but assume a fixed state space, limiting their ability to generalize to novel states. We consider the problem of safe reinforcement learning that must adapt to novel, potentially unsafe states. Our proposed approach for context aware policy adaptation leverages foundation models as a contextual representation that enables the agent to align novel observations with similar experience. We demonstrate empirically that our approach is able to generalize across different types of novelty that may include dangerous as well as safe states. We also show performance and safety are robust even when multiple types of novelty are introduced.
Submission Number: 6
Loading