Scientists from Google Research, who now work at Google DeepMind, have created a quadruped robot benchmark called Barkour. The purpose of the benchmark is to comprehensively measure robot agility and mobility and to develop locomotion controls for robots that can move quickly, in a controlled manner and in a variety of ways. The team took inspiration from agility competitions involving dogs, which have to complete an obstacle course.
Like dogs in these competitions, the four-legged robots have to master a challenging obstacle course that includes moving in different directions, traversing uneven terrain and jumping over obstacles. The 5 m × 5 m course includes four obstacles, but it can be adapted for larger areas and include a variable number of obstacles and course configurations.
The robot starts on a launch table, weaves its way through a series of poles and scales an A-frame. After that, it has to do a 0.5 m jump and climb onto a target table. The Barkour Judging System includes a target time to clear a single obstacle and a time for the entire course. The rating is between 0 and 1 for mastering the entire course. If a robot does not complete an obstacle in the target time, it will receive a time penalty.
The Google scientists used Reinforcement Learning (RL) to train the robot’s individual, specialized locomotion skills in a simulation. The robot learned walking, climbing and jumping strategies. Then they trained the transitions between the individual obstacles. In addition, they equipped the robot with the ability to recover itself, for example when stumbling on an obstacle. The robot trained in this way achieved agility values between 0.75 and 0.9.
The scientists believe that the Barkour benchmark is suitable for quantifying the agility of multi-legged robots at the animal level. The benchmark is demanding and can be easily adjusted.
