Barkour

A quadruped-robot agility benchmark from Google, modeled on dog agility trials, for measuring animal-level agility in a repeatable way.

Barkour is a quadruped-robot agility benchmark released by the Google Research team (later folded into Google DeepMind) in May 2023, its name blending “bark” and “parkour.” Modeled on canine agility trials, it sets up weave poles, an A-frame ramp, a 0.5-meter broad jump, and a finish-line table in a 5-meter-by-5-meter course, scoring 0 to 1 based on whether each obstacle is cleared and whether the time comes close to that of a small dog (about 10 seconds). The team first trained specialized skills — walking, climbing, jumping — separately with reinforcement learning, then used teacher-student distillation to combine them into a single Transformer-based generalist locomotion policy, which completed the course on the team's own quadruped robot in about 20 seconds, roughly half a small dog's speed. Its contribution is giving “animal-level agility” a quantifiable, reproducible way to compare systems.

ExampleThe team's own quadruped robot typically completes the Barkour course in about 20 seconds, versus about 10 seconds for a small dog.

Also called

Barkour: Benchmarking Animal-level Agility with Quadruped Robots

Related

Quadruped Robot · Legged Locomotion · Parkour · Benchmark · Teacher-Student Distillation · RL-based Locomotion Control

Sources

Barkour: Benchmarking Animal-level Agility with Quadruped Robots (arXiv 2305.14654)
Barkour: Benchmarking animal-level agility with quadruped robots (Google Research Blog)

As of

2023-05