Argo AI has released the technical guidelines it applies to facilitate safe interactions between autonomous vehicles and cyclists. The guidelines were created in collaboration with The League of American Bicyclists and are intended as a foundation for innovation and improvement among companies developing self-driving technologies.
To understand concerns among cyclists when sharing the road, Argo set out to collaborate and engage with the cycling community. The League of American Bicyclists provided consultation to inform Argo of common cyclist behaviors and typical interactions with vehicles. Together they outlined six technical guidelines for the manner in which a self-driving system should accurately detect cyclists, predict cyclist behavior, and drive in a consistent way to effectively and safely share the road.
Due to the unique behaviors of cyclists that distinguish them from scooter users or pedestrians, a self-driving system (SDS) should designate cyclists as a core object representation within its perception system to detect cyclists accurately. By treating cyclists as a distinct class and labeling a diverse set of bicycle imagery, a SDS detects them in a variety of positions, orientations, viewpoints, and speeds. It should also account for the different shapes and sizes of bikes as well as different types of riders.
An advanced understanding of potential movement patterns is necessary to predicting the intentions of a cyclist and prepare the self-driving vehicle’s actions. A cyclist may lane split, yield at stop signs, walk a bicycle, or make quick lateral movements to avoid obstacles on the road. A SDS should utilize specialized, cyclist-specific, motion forecasting models that account for a variety of behaviors. As a result, when the self-driving vehicle encounters a cyclist, it generates multiple possible trajectories – capturing the potential options of a cyclist’s path – enabling the SDS to better predict and respond to the cyclist’s actions.
A SDS should use high- definition 3D maps that incorporate details about cycling infrastructure, including the location of dedicated bike lanes, and all local and state cycling laws, to ensure its SDS is compliant. Accounting for this infrastructure enables the SDS to anticipate cyclists, and maintain a safe distance between the self-driving vehicle and the bike lane. When driving alongside a bike lane, the system will consider the higher potential for encountering a cyclist and common cyclist behavior. This behavior may see the cyclist merge into traffic to avoid parked cars blocking a bike lane, or treating a red light as a stop sign in some states.
Developers of self-driving technology should strive for the technology to operate in a naturalistic way so that the intentions of autonomous vehicles are understood by other road users. In the presence of nearby cyclists, or when passing or driving behind cyclists, a SDS should target conservative and appropriate speeds in accordance with local speed limits. It should also target margins that are equal to, or greater than, local laws and only pass a cyclist when it can maintain those margins and speeds for the entire maneuver.
In situations where a cyclist encroaches on a self-driving vehicle, when lane splitting between cars during stopped traffic for example, the vehicle should minimize the use of actions which further reduce the margin or risk unsettling the cyclist’s expectations. The SDS should also maintain adequate following distances – so that if a cyclist happens to fall, the self-driving vehicle has sufficient opportunity to maneuver or brake. Self-driving vehicles should also provide clear indications of intentions, including using turn signals and adjusting vehicle position in-lane when preparing to pass, merge lanes, or turn.
A self-driving system should account for uncertainty in cyclists’ intent, direction, and speed. For instance, reducing vehicle speed when a cyclist is traveling in the opposite direction of the vehicle in the same lane. When there is uncertainty, the SDS should lower the vehicle’s speed and, when possible, increase the margin of distance to increase the time and distance between the vehicle and the cyclist and drive accordingly.
The key to developing safe and robust autonomy software is thorough testing. Developers of self-driving technology should be committed to continuous virtual and physical testing of its self-driving system with a focus on cyclist safety across all phases of development, including virtual and physical testing.
The creation and simulation of real-life scenarios in the virtual world to safely test a wide variety of scenarios. A virtual testing program should be made up of three main test methodologies (simulation, resimulation, and playforward) to test an exhaustive permutation of autonomous vehicle and cyclist interactions on a daily basis. These scenarios should capture both varying vehicle and cyclist behavior, as well as changes in social context, road structure, and visibility.
Physical testing includes testing on closed courses followed by public roads. Testing on a test track validates simulation and ensures the technology behaves in the real world as it did in the virtual one. Scenarios tested should include interactions that are likely to occur on both public roads and rare situations. Fleet testing on public roads in multiple cities exposes the technology to a diverse variety of urban environments to learn about cyclist behaviors and validate that the self-driving system works as intended.
The development and publication of these guidelines are intended for adoption as industry best practices, promoting special consideration of cyclist behavior and interactions. Argo and The League have said that they encourage them to be used by all self-driving technology developers to build trust in self-driving technologies as testing and deployments expand.
These guidelines build upon the six principles Argo outlined last year for the development of a self-driving system that prioritizes safe interactions with vulnerable road users, through which it set out to contribute to an environment of collaboration, engagement, and education with community advocacy groups.