Adaptation of Quadruped Robot Locomotion with Meta-Learning

Abstract

Animals have remarkable abilities to adapt locomotion to different terrains and tasks. However, robots trained by means of reinforcement learning are typically able to solve only a single task and a transferred policy is usually inferior to that trained from scratch. In this work, we demonstrate that meta-reinforcement learning can be used to successfully train a robot capable to solve a wide range of locomotion tasks. The performance of the meta-trained robot is similar to that of a robot that is trained on a single task.

14 pages, 6 figures

References in corpus (11)

• Solving Rubik's Cube with a Robot Hand
• RL${}^2$: Fast Reinforcement Learning via Slow Reinforcement Learning
• Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Variables
• Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World
• Deep Dynamics Models for Learning Dexterous Manipulation
• VariBAD: A Very Good Method for Bayes-Adaptive Deep RL via Meta-Learning
• Controlling Overestimation Bias with Truncated Mixture of Continuous Distributional Quantile Critics
• Active Domain Randomization
• Emergent Real-World Robotic Skills via Unsupervised Off-Policy Reinforcement Learning
• Auto-Tuned Sim-to-Real Transfer
• MELD: Meta-Reinforcement Learning from Images via Latent State Models