Meta Reinforcement Learning for Optimal Design of Legged Robots

The process of robot design is a complex task and the majority decisions are still based on human intuition or tedious manual tuning. A more informed way facing this computational methods where parameters concurrently optimized with corresponding controllers. Existing approaches, however, strongly influenced by predefined control rules motion templates cannot provide end-to-end solutions. In paper, we present optimization framework using model-free meta reinforcement learning, its application to optimizing kinematics actuator quadrupedal robots. We use learning train locomotion policy that can quickly adapt different designs. This used evaluate each instance during optimization. demonstrate robots designs track random velocity commands over various rough terrains. With controlled experiments, show achieves close-to-optimal performance for after adaptation. Lastly, compare our results against model-based baseline approach allows higher while not being constrained motions gait patterns.