Building the Code to Help Autonomous Vehicles Learn
Training wheels are a safety measure to help a child balance when learning to ride a bike. Inspired by this observation, a group of scientists at the University of Michigan is developing algorithms that help autonomous vehicles learn through experiments to operate safely.
“When people learn how to do something, they either overcompensate for safety or take too much risk,” said Ilya Kolmanovsky, professor of Aerospace Engineering at UMich. “Using our method, we can provide a vehicle with the capability to learn a more precise control at an earlier stage to prevent damage and ensure safety and efficiency.”
To tackle these challenges, Kolmanovsky and Anouck Girard co-led a research team that explored the question of safety in autonomous vehicles from the two opposite perspectives. Starting from the system without any safety controls, a more ‘free-wheeling’ approach, they gradually tighten constraints on the system to ensure safety. They also put electronic ‘training wheels’ on the system and gradually adjust them to maximize the functionality. By approaching this problem from either end of the spectrum, they believe they have developed algorithms that converge on safety.
The researchers will present their work during the 2020 American Control Conference in Denver.
The algorithms they are creating are nimble and flexible and help the machines learn how to avoid situations that result in dangerous accidents. The researchers incorporated a control unit, called an explicit reference governor (ERG), to pre-filter commands. The ERG also works well with legacy hardware, limiting the need to update systems in a fleet. Most importantly, this component can learn. The ERG incorporates new information into the evaluation process to protect the system from future adverse conditions.
“If you are too cautious, you are not exploiting the full capability of the system,” Girard said. “The vehicles are learning to increase risk without breeching the safety limits.”
The team demonstrated the algorithm’s functionality through a series of simulations. One simulation examined how the algorithms help an autonomous vehicle avoid roll-over situations by learning to control the steering command. During the ‘free-wheeling’ scenario, the vehicle does not initially restrict the steering control and has a greater risk of roll over. In the ‘training wheels’ scenario, the steering command is held under much tighter control initially. Information obtained from each situation informs the next simulation as the vehicle learns to operate safely.
“What is exciting is that we can find approaches and algorithms for the system to learn autonomously and to guarantee safety without much knowledge about the system,” said Kolmanovsky. “We have come up with the very promising developments thus far.”
Kolmanovsky cautions that there is more work to be done on improving how quickly the vehicles learn.
K. Liu, N. Li, I. Kolmanovsky and A. Girard from the University of Michigan are working on a project, titled “Autonomously learning mobility limits” in collaboration with D. Rizzo from the US Army Combat Capabilities Development Command Ground Vehicle System Center and E. Garone from Université Libre de Bruxelles.
Stacy W. Kish