ABS

CMU's 2024 quadruped framework: sprint at high speed, switching to a learned collision-avoidance policy whenever a learned safety value says to.

ABS was developed by Guanya Shi and Changliu Liu's groups at Carnegie Mellon University with ETH Zurich, published at RSS 2024. Earlier quadruped obstacle-avoidance controllers, in the name of safety, usually stayed under 1 m/s; controllers built purely for agility didn't worry about collisions at all. ABS uses two policies: an agile policy for high-speed running between obstacles, and a recovery policy for emergency avoidance; a learned reach-avoid value network estimates in real time whether the current state is safe, decides when to switch between the two, and supplies the optimization target for the recovery policy. Obstacle information is compressed into distances along 11 rays, predicted by a network from a depth image. Each module is trained in Isaac Gym and deployed to a real Unitree Go1, running entirely on onboard sensing and compute, reaching a peak speed of 3.1 m/s. It's a representative example of combining control-theoretic safety guarantees with reinforcement-learning-based locomotion.

ExampleA Unitree Go1 sprints down a dim hallway at an average of 1.5 m/s and a peak of 2.5 m/s; when someone walks toward it or suddenly sticks out a leg to block it, a drop in the safety value triggers a switch to the recovery policy to dodge, then switches back to the agile policy to keep running once it's safe again.

Also called

Agile But Safe, Agile But Safe: Learning Collision-Free High-Speed Legged Locomotion

Related

Legged Locomotion · RL-based Locomotion Control · Obstacle Avoidance · Safe Reinforcement Learning · Hamilton-Jacobi Reachability Analysis · Unitree Go1

Sources

Agile But Safe: Learning Collision-Free High-Speed Legged Locomotion (arXiv 2401.17583)
ABS 项目页 (Chinese)
Robotics: Science and Systems XX (RSS 2024) 论文集 (Chinese)

As of

2024-05