EXTREME PARKOUR WITH LEGGED ROBOTS

Summary from the TLDR newsletter

This study is an attempt at developing robot parkour on a small low-cost robot with imprecise actuation and a single front-facing depth camera for perception. It shows how a single neural net policy operating directly from a camera image can overcome imprecise sensing and actuation to output highly precise control behavior end-to-end. The resulting robot is able to high jump on obstacles two times its height, long jump gaps two times its length, do handstands, run across tilted ramps, and more. A video of the robot performing different movements is available.