Learning Locomotion Skills Using DeepRL: Does the Choice of Action Space Matter?
This addresses the problem of optimizing action spaces for researchers and practitioners in robotics and reinforcement learning, but it is incremental as it builds on existing methods without introducing a new paradigm.
The study investigated how different action parameterizations (torques, muscle-activations, target joint angles, and target joint-angle velocities) affect learning difficulty and performance in deep reinforcement learning for locomotion skills, showing that higher-level parameterizations improve learning, robustness, and motion quality in gait-cycle imitation tasks.
The use of deep reinforcement learning allows for high-dimensional state descriptors, but little is known about how the choice of action representation impacts the learning difficulty and the resulting performance. We compare the impact of four different action parameterizations (torques, muscle-activations, target joint angles, and target joint-angle velocities) in terms of learning time, policy robustness, motion quality, and policy query rates. Our results are evaluated on a gait-cycle imitation task for multiple planar articulated figures and multiple gaits. We demonstrate that the local feedback provided by higher-level action parameterizations can significantly impact the learning, robustness, and quality of the resulting policies.