Argo AI shares tips for self-driving cars to safely share roads with cyclists
Cyclists can create unique challenges for human drivers. Their behavior is hard to predict, and they may change lanes or swerve around obstacles faster than other drivers. According to NHTSA (the National Highway Traffic Safety Administration), in 2020 the number of fatal crashes involving cyclists rose by 5% (846) compared with 2019.
Autonomous vehicles face the same challenges, so companies developing self-driving systems must take special care when it comes to cyclists. Although autonomous cars could reduce the number of road accidents and make roads safer, they first need to learn how to navigate the road as effectively as possible.
Argo AI, which is developing its own self-driving system, together with the League of American Bicyclists, published its best practices for building self-driving systems that work safely with cyclists.
“Argo AI and the League of American Bicyclists share a common goal: making streets safer for all road users. We appreciate Argo's proactive approach to research, development and testing for the safety of people outside vehicles. Roads have become significantly less safe for people outside vehicles over the past decade, and by addressing cyclist interaction now, Argo is demonstrating a commitment to the role of automated technology in reversing this deadly trend”
Here is a brief overview of Argo AI's six recommendations that companies should follow when developing self-driving systems.
"Learn about the technical recommendations developed by Argo together with the League of American Bicyclists to ensure safe interaction between autonomous vehicles and cyclists. These recommendations represent best industry practices for strengthening trust in self-driving technology and encouraging cooperation with the cycling community," Argo AI said.
Cyclists should be a unique class of objects
For a self-driving car to react to a cyclist on the road, it must first understand what cyclists are and how they move. Cyclists behave on the road differently from everyone else, so it is important that the self-driving system treats them as a primary object in its perception system.
By labeling a diverse set of bicycle images, an autonomous system can recognize a cyclist from any viewpoint, speed, position or bike orientation. Self-driving systems should also be able to recognize bicycles of different shapes and sizes, such as recumbent or electric ones.
Typical cyclist behavior should be expected
There are certain behaviors characteristic of cyclists that self-driving systems must be able to recognize and anticipate. Splitting lanes, yielding at stop signs, or walking a bike are behaviors that self-driving systems must recognize and respond to appropriately.
Self-driving systems should use specialized prediction models for cyclists. With these models, when a car meets a cyclist, it can predict many possible paths for them, making it easier to anticipate and respond to the cyclist's actions.
The following 2019 video shows how a Waymo autonomous vehicle approaches two cyclists and a vehicle blocking the bike lane. The self-driving system correctly predicts that the cyclists will pass the vehicle on the left and slows down to let them through.
Bike infrastructure and local laws must be mapped
Many cities and states have special laws for cyclists that a self-driving car should know about. For example, in some states cyclists are allowed to treat a red light as a stop sign. In addition, all bike lanes should be marked on the map for self-driving cars.
Knowing where bike lanes are will allow a self-driving car to anticipate more cyclists in those areas and closely monitor typical cyclist behavior.
Consistent and predictable control
The goal of self-driving cars is to replicate and improve on human driving. This means self-driving cars should behave with cyclists and other drivers on the road just as a human driver would. Self-driving cars should use turn signals wherever possible to help others understand their intentions.
Self-driving cars should also keep a greater distance from cyclists, just as a human driver would. They should take special care when passing cyclists.
Preparing for uncertain situations
It is inevitable that sometimes drivers and cyclists act unpredictably, and self-driving cars do not know how to react to them. In this case, the car should slow down and, where possible, increase the distance between the car and the cyclist. Self-driving systems should always account for uncertainty.
Scenarios involving cyclists must be constantly tested
Thorough testing is one of the most important aspects of developing a self-driving system. Testing both in simulation and in the real world is critical to building a system that can work safely with cyclists.
Simulation should be used to test real-world scenarios in a virtual environment to safely check a variety of situations. These scenarios should reflect the behavior of vehicles and cyclists, as well as changes in road structure and visibility. Real-world trials should be used to verify the simulation and ensure the technology performs the same as it did in testing.
