Argo AI Releases Standards for Self-Driving Car & Cyclist Interaction

Improving Autonomous Vehicle Interaction with Cyclists

Argo AI, a leading company in autonomous driving technology, has collaborated with the League of American Cyclists (LAB) to establish guidelines for how self-driving vehicles should perceive and interact with cyclists. This initiative aims to create an industry standard, particularly as autonomous vehicle technology transitions from testing phases to broader commercial deployment.

The World Health Organization estimates approximately 41,000 cyclist fatalities occur annually due to road traffic incidents. While autonomous vehicles are projected to significantly reduce collisions, this safety relies heavily on robust initial programming. Self-driving cars utilize extensive databases to identify objects and potential scenarios, and Argo’s guidelines prioritize training models to specifically recognize cyclists, cycling infrastructure, and relevant traffic regulations.

Peter Rander, president and co-founder of Argo AI, stated that developing these guidelines demonstrates a commitment to building public trust and ensuring a comfortable, safe experience for cyclists. He further encouraged other autonomous vehicle developers to adopt these practices to enhance trust among vulnerable road users.

Argo, currently conducting self-driving vehicle tests in the U.S. and Germany, engaged with the LAB community to understand typical cyclist behaviors and interactions with vehicles. This collaboration resulted in six key technical guidelines focused on cyclist detection, behavior prediction, and consistent driving protocols.

Cyclists as a Unique Object Class

To facilitate accurate identification, cyclists should be categorized as a distinct object class within the self-driving system. This involves training the system with a diverse range of bicycle imagery, encompassing various positions, orientations, viewpoints, and speeds.

According to Argo, this approach also accounts for the differing shapes and sizes of bicycles and their riders. A dedicated object representation is crucial for the accurate detection of cyclists by the self-driving system (SDS).

Predicting Cyclist Behavior

Cyclists often exhibit unpredictable actions, such as lane splitting, dismounting, or making sudden maneuvers to avoid road hazards. A sophisticated self-driving system must anticipate these behaviors and react appropriately.

The system should employ specialized motion forecasting models tailored to cyclist behavior. This includes generating multiple potential trajectories to account for various possible paths, enabling the SDS to better predict and respond to a cyclist’s actions.

Mapping Infrastructure and Laws

Self-driving systems rely on high-definition 3D maps to interpret their surroundings. These maps should incorporate cycling infrastructure and local cycling laws. This allows the system to anticipate cyclist movements, such as merging to avoid obstacles or navigating intersections, and maintain a safe distance from bike lanes.

Consistent and Safe Interactions

Autonomous vehicle technology should operate predictably, ensuring cyclists can readily understand the vehicle’s intentions. This includes utilizing turn signals and adjusting vehicle positioning within the lane when preparing to pass, merge, or turn.

Argo emphasizes maintaining conservative speeds, adhering to local speed limits, and ensuring sufficient margins of safety. Passing a cyclist should only occur when these margins and speeds can be consistently maintained throughout the maneuver. The system should also provide ample space for cyclists in case of a fall, allowing for evasive action.

Preparing for Uncertainty

Self-driving systems must account for uncertainty regarding a cyclist’s intent, direction, and speed. For example, when encountering a cyclist traveling in the opposing lane, the vehicle should be trained to reduce speed.

Generally, the system should lower vehicle speed and increase the distance between the vehicle and cyclist in uncertain situations. This proactive approach to safety is becoming increasingly standard within the autonomous vehicle development community.

Continuous Testing of Cycling Scenarios

Rigorous testing is essential to validate the safety of autonomous vehicles. Argo and LAB recommend ongoing virtual and physical testing specifically focused on cyclist interactions.

Virtual testing should employ simulation, resimulation, and playforward methodologies to analyze a wide range of vehicle-cyclist interactions. These scenarios should encompass varying behaviors, road conditions, and visibility levels. Physical testing, conducted on closed courses and public roads, validates simulation results and ensures real-world performance.

Building Public Trust and Ensuring Safety

Gaining public acceptance is a significant challenge for the widespread adoption of autonomous vehicles. Recent surveys indicate that nearly half of respondents believe AVs are less safe than human-driven cars.

Demonstrating safety for all road users is crucial, but equally important is fostering public confidence in autonomous vehicle technology. Standardizing safety practices across the industry may be a key step toward achieving this goal.