Infer and Adapt
A demonstration tells us how a robot moved. Can it also tell us what made that motion useful? Infer and Adapt studies this question for bipedal walking on complex terrain.
Learn the reward, then the policy
We use inverse reinforcement learning to infer reward functions from expert locomotion demonstrations. The learned nonlinear rewards offer a way to examine the expert’s walking strategy. We then train locomotion policies using those rewards and evaluate their behavior on terrain outside the training setting.
The experiments show that policies trained with inferred rewards can improve walking on unseen terrain. The paper provides the algorithms, reward analysis, and experimental comparisons.
Publication
Related work
Learn to Teach studies the complementary problem of training a locomotion controller efficiently with privileged simulation information. Both papers are collected in the Humanoid Locomotion project.