Boston Dynamics brings its legged robots to the FIFA World Cup
AI News Desk
·
The Robot Report
··
7 min read
Atlas demonstrates the deceptive Ghost Rabona kick.
Atlas demonstrates the deceptive Ghost Rabona kick. Source: Hyundai
The FIFA World Cup is not only a showcase for global athletic competition and camaraderie; it’s also an opportunity for leading legged robots to strut their stuff. Boston Dynamics Inc. has taught its Atlas humanoid to kick a soccer ball while its Spot robots support security operations at tournament venues.
Hyundai Motor Co., which acquired Boston Dynamics in 2021, has expanded its 27-year partnership with FIFA World Cup. While other humanoid robots have participated in the RoboCup league, Hyundai is leaving the competition to humans while demonstrating how its robots can learn, move, and do actual work.
Alberto Rodriguez , director of robot behavior, and Merry Frayne , senior director of product for Spot at Boston Dynamics, spoke with The Robot Report about the company ‘s involvement with this year’s World Cup.
How did this partnership with FIFA come about?
Rodriguez: We’re always looking for artistic expression to drive the technology forward, from dancing to parkour. The World Cup was too good of an opportunity to pass on.
Hyundai came to us and proposed a focused campaign based on soccer. On the technical side, we were extremely excited, because soccer has extreme behaviors, and the time was right to challenge the new generation of Atlas to do things like imitate the iconic trick shots of leading players.
For the “ School of Football ” campaign, we developed the “Ghost Rabona” kick. We’ve done lots of athletic maneuvers — jumps and flips — but they’re not as complex as the interactions with an object [the ball] and the ground. Practicing the motions for quick footwork is more complex than a backflip.
How did you teach Atlas these complex motions?
Rodriguez: The technologies to capture motion and design behaviors today make it easier to explore a superhuman range of motion. For the most complex trick shots, we can use motion capture or mocap. But that’s constrained to the human form factor.
In simulation , we can optimize that motion without constraints and find the most robust way to do things within Atlas’ constraints and with a large number of randomizations. We learned how to push the soccer ball, position the robot’s arms, and deal with the coefficient of friction with the ground.
We’re really conscious of the decision to move to reinforcement learning [RL] for robotic control, which makes it easier to solve problems.
Rodriguez: About a year ago, we had the best implementation in the world of predictive control for humanoid balance. We had invested many years in that model based on earlier work in athletic behavior.
We started from scratch with RL. Atlas moving a fridge [see video below] would have been difficult to do with classical model-based techniques. Reinforcement learning makes it possible to go beyond what you can demonstrate.
We put a lot of effort into designing the robot to be very simple from a mechanical perspective. It’s all rotary actuators rather than linear, making it easier to simulate and to manufacture.
We wanted the capacity on the software side to learn in simulation and for it to just work on the robot. The same techniques we used to get Atlas to carry a fridge enable mobile manipulation for carrying a box in the warehouse.
How did his enable the Ghost Rabona move?
Rodriguez: It was way faster to train with this technique. We put someone good at soccer in a mocap suit and studio.
We put that data into a simulation engine, and the policy came out only a few hours to a day after that.
Once we expand, we challenge ourselves to do something more complex. It could be to adjust a recipe for domain presentation so the behavior is more reliable. We can change something in the simulation if it’s not properly represented.
This could also be good for other things, and we might need to spend some time to improve the whole process.
Can you give an example of the challenges unique to playing football or soccer?
Rodriguez: A grass field is way more compliant than concrete, plywood, or many other surfaces. It was not properly represented in our simulation engine, so the algorithm had to be robust enough to deal with unexpected compliance. The robot’s foot can sink, slip, or get caught on turf.
The quick footwork and motion of the lower body challenged us. We had to improve how these control policies are deployed on the robot and reduce latency so it could react fast enough to not trip or slip.
Now, all of Atlas’ behaviors are informed by this. This work with Atlas will translate to industry, such as more robust stepping in different terrains.
We’re increasing public awareness of our robots’ capabilities, but it’s also important to be clear that they’re far from doing things that a real soccer player has to do, including interact with other players or safely run next to people — the technology isn’t there today.
How is Boston Dynamics’ Spot quadruped supporting FIFA security operations?
Rodriguez: Spot is being used in stadiums across the U.S. to support public safety, just as it’s used by other organizations in other instances, such as industrial inspection .
The robot is there as a tool, but it doesn’t have facial-recognition capabilities.
Frayne: Spot is a working dog — it’s not a photo op. The robot is walking around two stadiums, looking for hazards and suspicious packages.
Spot Cam is working as a second pair of eyes. There’s a big 360-degree camera on the front of the robot, plus 4k, high-def, thermal imaging, and an enhanced autonomy package with lidar.
Spot performed in America’s Got Talent last year. Where is the robot now working? Is it on the field or in the lockers?
Frayne: While we’re involved in serious commercial organizations and hardcore research, it’s also important to bring a sense of whimsy to help the public become more comfortable with robots.
For the World Cup, Spot is more on the perimeter, doing rounds when most people are gone. It can be around people, we can teach it to map and go around autonomously or through teleoperation . It can run in both modes.
Which World Cup venues are using Spot for security and inspection? How are they different from other implementations?
Frayne: There are two in Dallas and two at Citi Field.
Most of these robots are used behind closed doors at nuclear facilities or mines. They’re typically collecting data on how a facility is operating. Is anything out of the ordinary, like leaking or sagging pipes or overflowing trash cans?
How is security as a newer application for Spot?
Frayne: We haven’t had to modify it a lot for security as we move from industrial applications into the service area. For the public to start perceiving robots as a part of the world, doing useful work, we have to run the robot in a way that’s safe.
Who is running Spot at the World Cup venues?
Frayne: The security teams operate the robots themselves. We have development teams to help them get set up. We’ve had partners offering Spot robots as a security service for a while, and we’re building the use case.
It’s up to the security teams to decide how they want to use the robots. We’re a long ways from robots doing a full job, but they can now do useful tasks.
Spot is part of the security detail at FIFA World Cup stadiums. Source: Boston Dynamics
How are Boston Dynamics’ systems working with other robots and AI ?
Frayne: Spot’s routes go both indoors and outdoors. We find that applications for legged robots are complementary with wheeled and flying robots or drones.
With new AI tools that you can apply to data that robots like Spot are collecting, you can get richer understanding of the environment or communication with people. Stadiums are facilities that are built for people, which opens up a world of interesting opportunities.
Speaking of those opportunities, are there any particularly promising applications?
Frayne: We’re working with partners, including some that built custom payloads, building interesting momentum for our platform. For example, Path Robotics has partnered for welding.
We’re also starting to integrate interactive behaviors for circuit-breaker racking. It’s not just monitoring of data centers, but also construction of such complex environments.
We have a robust business with first responders, labs, and detecting and eliminating munitions. We’ll announce more major markets in the next few months.
The post Boston Dynamics brings its legged robots to the FIFA World Cup appeared first on The Robot Report .
