In this paper, the authors explore the challenge of getting robots to perform parkour, a highly dynamic and complex task that humans can accomplish with their precise coordination and movement. Unlike humans, robots have traditionally been limited to controlled settings like labs, where perception, actuation, and control systems are carefully engineered to strict tolerances. However, the authors take a different approach, using a small low-cost robot with imprecise actuation and a single front-facing depth camera. Through the training of a neural net policy in simulation, they demonstrate that the robot can overcome imprecise sensing and actuation to perform impressive parkour maneuvers such as high jumps, long jumps, handstands, and running across tilted ramps. The robot is also able to generalize its abilities to novel obstacle courses with different properties. This research represents a significant advancement in the field of robotics and showcases the potential for robots to master complex physical tasks.
https://extreme-parkour.github.io/