In this video, we demonstrate how our fully-learned method enables robots to conquer challenging scenarios reminiscent of parkour challenges.

The paper introduces a hierarchical formulation that trains advanced locomotion skills for various obstacles, including walking, jumping, climbing, and crouching. A high-level policy is used to select and control these skills, allowing the robot to adapt its behavior based on the environment at hand. Furthermore, a perception module is trained to reconstruct obstacles from occluded and noisy sensory data, enhancing the robot's scene-understanding capabilities.

Unlike previous attempts, our method does not require expert demonstration, offline computation, prior knowledge of the environment, or explicit consideration of contacts. It achieves impressive results solely through training on simulated data.

Our real-world experiments showcase the successful transfer of these learned skills onto hardware.

For more information:
- Visit our project website at https://sites.google.com/leggedroboti...
- Read the paper: https://www.science.org/doi/10.1126/s....