Creating robots with agile locomotion capabilities similar to animals or humans has been a long-standing goal in the robotics community. Google researchers have been working on this for years, but there is no widely accepted benchmark to measure robot agility. In “Barkour: Benchmarking Animal-level Agility with Quadruped Robots,” we introduce the Barkour agility benchmark for quadruped robots and a Transformer-based locomotion policy.
The Barkour benchmark is inspired by dog agility competitions and requires a robot to display various skills like moving in different directions, traversing uneven terrain, and jumping over obstacles within a limited timeframe. This benchmark encourages researchers to develop controllers that are both fast and versatile. We invited people to try the obstacle course, and while small dogs complete it in 10 seconds, our robot takes around 20 seconds.
The Barkour scoring system assigns a score between 0 and 1 based on the robot’s performance in traversing the obstacles within the given time. The robot receives penalties for skipping obstacles or moving too slowly. Our course includes weave poles, an A-frame, a broad jump, and pause tables. The scoring mechanism can be easily modified according to different obstacles or configurations.
To train the robot for the Barkour benchmark, we use a student-teacher framework combined with a sim-to-real approach. First, we train specialist locomotion skills for individual obstacles. Then, we train a single policy that performs all skills and transitions between them. We use simulation rollouts to create a dataset for each skill, and this dataset is then distilled into a single generalist locomotion policy.
During deployment, the robot is controlled based on its surroundings, velocity commands, and on-board sensory information. We also train a recovery policy to quickly get the robot back on its feet in case of failure during an obstacle.
We evaluate the performance of our locomotion policy using custom-built quadruped robots and show that it performs well in the real world. We find that failures are possible but can be handled with the recovery policy. Overall, our policies achieve similar performance and exhibit smooth transitions between behaviors and gaits.
We believe that the Barkour benchmark is a challenging and customizable way to measure progress in achieving animal-level agility in robots. It provides a basis for further research and development in the field of legged robotics.