Google-DeepMind's Barkour project is pushing the boundaries of robotics with robots that can match or even surpass the agility of real dogs. The good news is that the hardware has been open sourced.  

Meet Barky and friends, the agile quadruped robots which have been designed so that DeepMind can study "animal-level agility". But don't just watch the video and read the paper, you can download everything you need to replicate the hardware from the Google-Deepmind Barkour Robot repo on GitHub. The  repository contains the design assets - CAD, PCBA designs, assembly instructions and so on - necessary to build your own Barky licensed under an Apache 2 license. However while core software (firmware and low level code) is provided there's no higher level AI code provided - although it has been made easier to develop your own.

While we have seen successive generations of robot quadrupeds over the years, there had been little by way of benchmark tests to compare these robots' progress. To remedy this, starting in 2022, Google-DeepMind researchers developed an agility course which they called "Barkour". Drawing on real world dog agility competitions, it includes obstacles such as a pause table, weave poles, an A-frame climb, and a jump and uses a time-based scoring mechanism.

To set strong baselines, Google-Deepmind researchers devised two methods for tackling the benchmark. In the first approach, they trained specialist locomotion skills using reinforcement learning methods and combine them with a high-level navigation controller. In the second approach, they distilled the specialist skills into a Transformer-based generalist locomotion policy, that can handle various terrains and adjust the robot's gait based on the perceived environment and robot states.

They then tried it out with a custom-built quadruped robot, nicknamed "Barky" that was able to complete the course at half the speed of a dog - as demonstrated by training real dogs to navigate the course.: 

Full details are included in the report "Barkour: Benchmarking Animal-level Agility with Quadruped Robots" and summarized in this video: 

Barkour vb is also a recent four-footed addition of the MuJoCo menagerie. a collection of high quality, open sourced robot models for the MuJoCo physics engine, curated by Google DeepMind where it joins a line up of quadruped simulations including Unitree Go, Anybotics ANYmal and Boston Dynamics Spot.

More Information

Barkour on GitHub

MuJoCo Menagerie on Github

"Barkour: Benchmarking Animal-level Agility with Quadruped Robots" on arxiv

Related Articles

Quadrupedal Parkour

ANYmal Evolution and Locomotion

A Tale of Three (Robot) Dogs

Should We Beware The Unitree Go2

To be informed about new articles on I Programmer, sign up for our weekly newsletter, subscribe to the RSS feed and follow us on Twitter, Facebook or Linkedin.

Comments

Make a Comment or View Existing Comments Using Disqus

or email your comment to: comments@i-programmer.info