Terrain-Adaptive Locomotion Skills Using Deep Reinforcement Learning

摘要

Reinforcement learning offers a promising methodology for developingrnskills for simulated characters, but typically requires workingrnwith sparse hand-crafted features. Building on recent progress inrndeep reinforcement learning (DeepRL), we introduce a mixture ofrnactor-critic experts (MACE) approach that learns terrain-adaptiverndynamic locomotion skills using high-dimensional state and terrainrndescriptions as input, and parameterized leaps or steps as outputrnactions. MACE learns more quickly than a single actor-critic approachrnand results in actor-critic experts that exhibit specialization.rnAdditional elements of our solution that contribute towards efficientrnlearning include Boltzmann exploration and the use of initial actorrnbiases to encourage specialization. Results are demonstrated forrnmultiple planar characters and terrain classes.