By subscribing, you agree to our Terms of Use and Policies You may unsubscribe at any time.
In the world of robotics, quadrupedal robots are becoming increasingly impressive with their abilities and tricks. However, comparing these robots and evaluating their capabilities is challenging due to the absence of standardized metrics.
To address this issue, a team of research scientists at Google has come up with an innovative solution: robot obstacle courses inspired by dog agility competitions. This new approach, known as Barkour, aims to establish a benchmark for assessing the agility and mobility of quadruped robots, a blog post said.
The Barkour agility benchmark for quadruped robots is inspired by dog agility competitions. It requires robots to demonstrate a range of skills within a limited timeframe. This includes moving in different directions, navigating uneven terrains, and jumping over obstacles.
See AlsoThe benchmark aims to inspire researchers to develop robots that are not only fast but also agile, versatile, and controllable by offering a challenging and diversified obstacle course. Since the performance metric is based on the capabilities of real dogs, it provides a precise measure of how well robots perform compared to their animal counterparts.
Robots in Barkour are graded from 0 to 1 according on how well they navigate obstacles. Small dog agility establishes a target time and assigns consequences for skipping or failing obstacles as well as for moving slowly. In a 5x5m rectangle, the course has four different obstacles, including weaving, climbing, jumping, and stepping.
The choice of these obstacles ensures a diverse set of skills is tested while keeping the course within a manageable size. The Barkour benchmark is customizable and can be adapted to larger course areas with different configurations, similar to real dog agility competitions.
A learning approach is employed to achieve agile locomotion skills in the Barkour benchmark. Initially, individual specialist locomotion skills are trained for each obstacle using reinforcement learning methods. These skills include walking, slope climbing, and jumping.
By training the robot to master these specialized skills, it becomes equipped to tackle the various challenges presented by the Barkour course.
Next, a single policy called the locomotion transformer policy is trained to perform all the skills and transitions between them. This policy is based on a Transformer-based architecture. It enables the robot to adjust its gait based on the environment and its own state.
During deployment, the locomotion transformer policy is coupled with a navigation controller that generates velocity commands. These commands are determined by the robot's position and sensory data.
To ensure robustness, a recovery policy is trained to quickly restore the robot's stability in case of failures during the course. This allows the robot to continue the course without human intervention. The combination of specialist locomotion skills, the locomotion transformer policy, and the recovery policy enables the robot to tackle the Barkour course with agility and adaptability.
