<ul><li>Continual Learning (CL) is a method that enables agents to learn a sequence of tasks and use past knowledge for future learning.</li><li>Existing CL methods assume static agent capabilities in dynamic environments, not reflecting real-world scenarios.</li><li>A new problem, Continual Learning with Dynamic Capabilities (CL-DC), challenges agents to generalize policies across different action spaces.</li><li>The proposed Action-Adaptive Continual Learning framework (AACL) addresses this challenge by decoupling the agent's policy from specific action spaces and adapting action representations for new spaces, outperforming existing methods.</li></ul>

Action-Adaptive Continual Learning: Enabling Policy Generalization under Dynamic Action Spaces

Discover more