<ul><li>Skill-based reinforcement learning (SBRL) aims to quickly adapt in environments with sparse rewards by pretraining a skill-conditioned policy.</li><li>A new method called AMPED has been proposed to balance both exploration and skill diversification in SBRL.</li><li>AMPED introduces a gradient surgery technique during skill pretraining to manage conflicts between exploration and skill diversity objectives.</li><li>AMPED also includes a skill selector module for choosing suitable skills for specific tasks, outperforming SBRL baselines in various benchmarks.</li></ul>

AMPED: Adaptive Multi-objective Projection for balancing Exploration and skill Diversification

Discover more