Motion Policy for Highly Dynamic Moves for Robots

Here is an approach that enables robots to perform unseen highly dynamic motions. This paper explores the idea of using a variational encoder to extract latent space encoding and train a conditional policy using the latent code. The approach allows a bipedal robot to execute highly dynamic moves. The below video from DisneyResearchHub shows how it works:

The robot can regain its balance when pushed.

[HT]