Google DeepMind Develops Dog-Agility Course To Test Four-Legged Robots

A course to test the agility of four-legged robots is now available. The team behind the training course designed it after the dogs' dexterity.

Google DeepMind Designs Dog-Agility Courses for Robots

To test the capabilities of four-legged robots, a team of researchers from Google DeepMind created a robot-agility course dubbed Barkour inspired by canine agility courses, IEEE Spectrum reported.

Dogs have been taught to deftly jump over hoops, climb slopes, and weave between poles since the 1970s to exhibit their agility. They have been competing in these events and must be fast, react quickly, and pay attention to detail to win ribbons.

Additionally, these programs establish standards for measuring agility across breeds, which Atil Iscen, a Google DeepMind scientist in Denver, claims is lacking in four-legged robots.

Despite significant advancements over the last ten years, including robots like MIT's Mini Cheetah and Boston Dynamics' Spot that demonstrated how animal-like robots can move, Iscen claims that the absence of defined tasks for these robots makes assessing their progress challenging.

Iscen claims that Barkour differs from earlier benchmarks created for legged robots in that it has a variety of obstacles that call for a mix of several actions, including accurate walking, climbing, and jumping. Additionally, their timing-based metric for rewarding faster behavior motivates scientists to push speed limits while upholding accuracy and various motion standards.

The Barkour course was 25 meters square instead of the up to 743 meters square used for typical courses, which was the size of the agility course that Iscen designed.

Iscen and colleagues chose the pause table, weave poles, ascending an A-frame, and a jump from standard dog-agility courses for their smaller-sized agility course (the Barkour course was 25 meters squared instead of up to 743 square meters used for traditional classes).

They chose these challenges to test many facets of agility, such as speed, acceleration, and balance. The course can also be customized by extending the course to include additional sorts of obstacles within a greater region.

Robots competing on this course lose points for failing or missing obstacles and for going over the allotted time of approximately 11 seconds. It is similar to dog-agility competitions. The DeepMind team created two different simulations to gauge the difficulty of their course.

Robots competing on this course lose points if they fail or miss an obstacle, go over the allotted time of about 11 seconds, or both. It is similar to dog-agility competitions. The DeepMind team created two distinct learning strategies for their course to determine how challenging it was: a specialist strategy that trained on each type of skill required for the course, for example, jumping or slope climbing- and a generalist strategy trained by studying simulations run using the specialist strategy.

ALSO READ: Hospital Evolution: Autonomous Surgical Robots, 3D Organs, Smart Toilet That Detects Diseases, More Are Now Available

Real Dogs Exhibit Better Agility Than Four-Legged Robots

When the team unleashed the four-legged robots they had trained in these two distinct ways onto the course, they discovered that the robots taught with the specialist technique had barely beaten the robots trained with the generic strategy. The generalists took closer to 27 seconds to finish the course, while the specialists took about 25 seconds. However, two little dogs-a Pomeranian/Chihuahua mix and a Dachshund-who finished the course in less than 10 seconds, outperformed robots trained with both methods and went over the course time restriction.

This benchmark shows that there is still a significant agility difference between robots and their animal counterparts, according to the research team.

The team claims that although the robots' performance may not have lived up to expectations, this is good because it shows that there is still room for development and progress.

Iscen believes that the Barkour course's simplicity of replication will make it a desirable benchmark in the future for use in the field.

He added that they proactively considered the benchmark's reproducibility and kept the material costs and environmental footprint low.

If other research teams interested in the project get in touch with them, Iscen said they would be pleased to share the lessons learned about developing it and would love to see Barkour setups appear in other labs. They hope more labs will adopt this benchmark so the entire community can work together to solve this difficult issue.

Iscen claims that the DeepMind team is also eager to investigate how human partners play a part in dog agility training in their upcoming work.

Agility competitions initially seem to be only focused on the dog's performance, according to Iscen. However, according to him, a lot depends on the brief information exchanges between the dog and its handler. Thus, they are excited to investigate human-robot interactions in this setting, such as how a handler may cooperate with a legged robot to assist it quickly through a novel obstacle course.

The DeepMind's Barkour course was published on the arXiv preprint server in May.