CMU's Roborace Team Launches Virtual, Autonomous Racing Challenge

A virtual, autonomous racing challenge launching this week will enable aspiring drivers to head to the track without leaving their computer.

The Learn-to-Race Autonomous Racing Virtual Challenge started Monday, Dec. 6. Competitors use the Learn-to-Race environment to teach an artificially intelligent agent how to race. The challenge is coupled with a workshop on Safe Learning for Autonomous Driving, which is accepting research paper submissions.

"We want people to use Learn-to-Race, make improvements to the environment, push it to the limit and create an agent that could run on a track," said James Herman, a CMU alumnus who wrote the Learn-to-Race framework and is part of the team launching the challenge. "Hopefully, people will have fun with it and come up with creative ideas."

Herman graduated this past May from the Master of Computational Data Science (MCDS) program in the Language Technologies Institute (LTI) at CMU's School of Computer Science. He now works in New York as a machine learning engineer at Intellimize.

What drew Herman to autonomous racing had nothing to do with cars or going fast. During Herman's first month at CMU, Anirudh Koul, also an MCDS alumnus, gave a presentation that captured his attention, and Herman decided he wanted to work with Koul. That next semester, Koul announced he was sponsoring a capstone project with Siddha Ganju, another MCDS alumna, to work on autonomous racing. Herman was interested.

That capstone project led to Herman proposing that CMU enter the Roborace Challenge, a global autonomous racing competition. CMU's team became the first from North America to enter, placing second and third in events during its first season.

Learn-to-Race grew out of that first season as a broader initiative to support research in achieving the safety benefits of autonomous vehicles through simulated racing environments. Jonathan Francis, another mentor for the Roborace project who is a Ph.D. candidate in the LTI and a research scientist at Bosch Research Pittsburgh, leads the initiative with Bingqing Chen, a Ph.D. candidate in CMU's College of Engineering.

Learn-to-Race uses reinforcement learning to teach agents how to race. The more an agent races, the better it gets as it explores the track and ways it can go faster. Learn-to-Race incentivizes the agent to do the latter by telling it that a good job means lowering the time it takes to travel from one point on the circuit to another. The environment also discourages the agent from performing unsafe actions by penalizing it for committing safety infractions, such as leaving the drivable area. The agent then, lap after lap, learns what actions it needs to take to go faster, maximize its progress and ensure safe exploration.

"The agents that we train learn from scratch," Herman said. "They don't know anything about racing. They don't know anything about driving."

The Learn-to-Race framework is built around the Arrival Autonomous Racing Simulator, which gives competitors access to models of real-world racetracks, like Thruxton Circuit in the United Kingdom and the North Road Course at the Las Vegas Motor Speedway.

"As kids, we all have felt the thrill of high-speed racing video games. MCDS has taken it to the next level by creating a reinforcement learning agent that learns to race in a controllable and realistic-looking environment," said Ganju, who works on self-driving cars and health care at NVIDIA and continues to mentor women in SCS. "We hope the simulator and Learn-to-Race environment make reinforcement learning much more accessible so anyone in the world can compete."

AIcrowd organized the competition and is hosting it for free.

"AIcrowd is excited to organize this competition. We are very happy to provide our platform at no cost for this competition as a part of our research initiatives at AIcrowd Research," said Sharada Mohanty, the company's CEO and co-founder.

Support from Amazon Web Services opens the competition to anyone for free. Herman said participants need only a basic understanding of coding to access the framework and simulator and start racing.

"AWS is proud to support Learn-to-Race by helping more aspiring developers off the machine learning starting line through the hands-on fun of racing autonomous vehicles," said Sahika Genc, principal scientist at AWS. "AWS is committed to providing access to educational resources to prepare students and professionals alike for artificial intelligence and machine learning career opportunities."

While autonomy will not replace a person's drive to go faster, it can help racers and everyday drivers follow speed limits, stay safe and maximize efficiency. The machinery used in the top echelons of motorsport becomes more and more advanced each year, and teams run hundreds of thousands of laps on simulators before a race. Acceleration, braking, steering angle, fuel usage, tire degradation and other factors can be precisely predicted and prescribed, leaving only how a driver will execute up to chance. Autonomy can help maximize the execution.

As the technology and ideas used in autonomous racing filter into race cars, they will certainly trickle down into road cars, improving current autonomous vehicles, their safety and the public's trust.

"If someone sees an autonomous vehicle safely race around a circuit, it could increase their confidence in an AV safely navigating their city streets," Herman said. "And if an autonomous agent could beat a human in a live race, then it could get interesting."

Learn more about the Learn-to-Race Challenge and follow it on Twitter.