The Thunderhill Raceway in California’s Sacramento Valley is soaked in the oil of motorsport history.
It is home to the longest automobile race in the United States, known as the 25 Hours of Thunderhill, and last September, it was host to a very different kind of race: a race between man and machine.
In the distance, a motorbike driver took the curves of the course like any other professional rider. It was only up close that reality hit the spectator: the rider wasn’t human. It was a blue humanoid robot that looked like it had stepped straight out of a screenshot from the computer game Halo.
Halo is set in an interstellar war between humanity and a bunch of aliens called the Covenant; you get the picture. When this robot – Motobot – stops driving, you feel like it could climb off the bike and hunt you down as well – but it can’t, yet.
Motobot 2.0 is a fully autonomous motorcycle-riding robot that was specially designed to drive around a racetrack at high speed on a Yamaha YZF-R1M, the same kind of motorbike that racing legend Valentino Rossi rides. Human operators can specify how aggressively Motobot should race on a scale of zero to 100% and it does the rest. This process is roughly parallel to how a racing team might discuss strategy with a human rider. The bike itself looks like the classic modern aerodynamic racing bike that overtakes you on the motorway.
In September, the team that developed Motobot 2.0 achieved one of its goals when the robot successfully hit speeds exceeding 200km/h (124 mph) on the race track – 50 km/h faster than its predecessor Motobot 1.0. Unfortunately, or fortunately, depending on your point of view, this still fell short of the lap time of Valentino Rossi – which they were trying to beat – by about 30 seconds.
We had several major crashes, two were catastrophic – Hiroshi Saijou
Rossi is one of the most successful motorcycle racers of all time, with nine Grand Prix World Championships to his name. His best lap time on the course was a very quick 85 seconds. Motobot tried to break Rossi’s record a second time in October – and failed again.
“We had several major crashes, two were catastrophic,” says Hiroshi Saijou, CEO and managing director at Yamaha Motor Ventures & Laboratory Silicon Valley, and the lead on the Motobot project. “Nobody was injured, and both happened in safe, to human, controlled circumstances. Several times we've had low-speed incidents as we were developing and testing Motobot. This is all part of the price of pushing the boundaries of technology.”
Motobot started life in 2014 as a ‘moonshot’ project for Japanese motorcycle manufacturer Yamaha. Moonshots are those “it would be great if they worked but…” projects: ambitious, exploratory and ground-breaking but with little hope of short-term profitability.
Yamaha’s initial concept was a “humanoid robot that can ride a motorcycle autonomously” and the company teamed up with SRI International to achieve its vision. SRI, the Stanford Research Institute, as it was originally known, was founded in 1946 to be the cutting edge of innovation in Silicon Valley. The institute has been responsible for the development of such projects as Apple assistant Siri, the computer mouse and humanoid robots such as Proxi, which is designed to assist humans after a natural disaster. Back in 1966 they built the first mobile robot with the ability to perceive and reason about its surroundings.
“Why a motorbike?” ponders Hiroshi Saijou. “Because it is very difficult to do, and it had never been done before.”
“The significance of the 200km/h goal was that it requires extremely high-speed prevision computing. Calculations must take place at 1/1000 of a second – and a minor mistake would be amplified and impossible for Motobot to recover from.
“Most human riders do not have the experience to ride at this speed. So, we set this speed as a good enough target to show that Motobot’s abilities were superior to human’s. Beating Rossi would have been clear evidence that Motobot can perform beyond human capabilities.”
For Brian Foster, robotics engineer and Motobot project lead at SRI International, another goal of the project was to “learn what makes a great rider.
“How riders sense the limits of traction, optimise power output of a bike and recover from exceeding the bike’s limits without crashing,” he says. “Using an unmodified bike was key to this and set the playing field for evaluating the robot versus the human competition.”
This rule meant that the designers then had to deal with constraints on all sorts of things like geometry, the size of the actuators that control the robot’s movements, where sensors were placed and more factors that wouldn’t have been an issue in a purpose-built vehicle. Yes, Motobot was physically attached to the bike but its hand was still required to grip and twist the throttle.
On the other hand, the robot didn’t need to use cameras or lasers to navigate like an autonomous car would because it wasn’t a public road. It could use technologies like the simpler GPS and IMUs (inertial measurement unit), that are often used to control such things as drones and satellites.
There were, however, plenty of challenges for the engineers to face before the robot could ride a bike very fast around the track without crashing it.
“Our first big challenge was the balance controller,” says Foster. “Motobot had to be taught how to balance the bike at lean angles from zero to over 50 degrees, at speeds between 5km/h and over 200 km/h. And it had to be able to change bank angles rapidly and precisely. The control algorithms to do this were constantly refined as we got close to our final high-performance version.
“Similarly, the path-following algorithm had to work well at high-speed straights, sweeping turns, hairpin turns, strong acceleration, and strong deceleration. Developing a controller that was adaptive to such a wide range of extreme conditions was a huge challenge.
“From my perspective though, the biggest challenge was identifying performance limits without crashing,” Foster adds. “To improve the algorithms, we had to constantly push it to the limit to see where improvements were needed. If we pushed too hard, we could crash and lose everything. If we didn’t push hard enough, we wouldn’t learn enough, and our progress would be too slow. It was a constant risk balance exercise.”
We went back and forth discussing what the limits to the competition should be – Stephen Morfey
To try to reduce the risk, Foster and his team would bring Motobot and the bike into the lab, where they ran a very sophisticated simulation whereby the robot would apply brakes and shift gears as if it were racing on the track. The sensors would then feed the data back into the simulation hundreds of times per second.
“Ultimately, nothing perfectly replicates the real world, so we still needed a lot of track time and had to manage the risks that come with that,” Foster says.
Hiroshi Saijou thinks the “cost to learn” is the reason why we didn’t see any depressing headlines about AI beating another human world champion.
“The most significant one is the cost – not only money but time and resources - to learn,” he says. “AI for a board game, such as AlphaGo, can learn how to play and how to win pretty quickly since there is no risk of it getting damaged. I believe that there were millions of failures before it eventually won over a human champion.
“For Motobot, the learning cost is way more expensive and repairs take a long time. So, we needed to take extraordinary care each time we did a trial.”
Perhaps Motobot needs a jetpack to beat Rossi.
“We went back and forth discussing what the limits to the competition should be,” says Stephen Morfey, a roboticist and now director of Morfey Design, a robot design consultancy. He was the lead mechanical designer on the Motobot project in its first phase and worked on other humanoid robots for SRI International like the walking bot Durus. “Jet thrusters weren’t allowed, but it could be aerodynamically shaped. We decided that physically attaching Motobot to the bike wasn’t cheating because its hands had to grip the handlebars.”
“At the start of the project, controlling Motobot was like playing a video game,” he adds. “You set the speed and told it the direction you wanted it to go. By the end, after I had left, it was autonomous.”
It would have been much easier to beat Rossi, he thinks, by designing from scratch a very fast autonomous two-wheeled vehicle. “No, we didn’t beat Rossi. Why not? Because it is a hard problem,” he says. “There are hundreds of different variables that you must consider. In principle, you can get a robot to optimise all this stuff, but in practice, it is much harder.”
While the failure of Motobot to beat Rossi’s time may have dented the pride of the engineering team involved, important lessons were learned.
The future of Motobot, it seems, might be on two legs. Motobot is different from most humanoid robots because it doesn’t walk… yet. But future versions might be able to walk up to the bike and get on it.
A kind of retrofitted autonomy, applicable to modern day problems, may have been made possible through their research and experiments. For example, in coming years, developing nations could use humanoid robots like Motobot to operate the perfectly good tractors and diggers that would have been replaced with new and expensive autonomous versions.
SRI International is already working with Chilean mining company Enaex to develop a rather freaky-looking remotely operated robot called Robominer that has the head, two arms and torso of a humanoid robot on four wheels.
Would Hiroshi Saijou classify Motobot a success? “It is still on the way. We have learned a lot in the last three years and will use this knowledge in our products in the future. It has huge potential for us to get real success in our business,” he says. “What we learned is so unique that it would have been hard to get without Motobot. We are actively working on Motobot 3.0. Please stay tuned.”
Rossi, you have been warned.
Join 800,000+ Future fans by liking us on Facebook, or follow us on Twitter.