Self-driving cars have the potential to significantly reduce traffic accidents caused by human error, which account for over 90% of all accidents according to the National Highway Traffic Safety Administration. For autonomous vehicles to be deployed safely on public roads, robust safety measures need to be in place. Vehicle manufacturers and researchers are taking safety very seriously and implementing redundant systems to minimize risks.
One of the most important safety aspects of self-driving car design is sensors and perception. Autonomous vehicles use cameras, lidar, radar and ultrasonic sensors to perceive the environment around the vehicle in all directions at once. These sensors provide a 360 degree awareness that humans cannot match. Relying on any single sensor could potentially lead to accidents if it fails or is disrupted. Therefore, multiple redundant sensors are used so that the vehicle can still drive safely even if one or more sensors experience an outage. For example, a vehicle may use four long range lidars, six cameras, twelve short-range ultrasonic sensors and four radars to observe the surroundings. The data from these diverse sensors is cross-checked against each other in real-time to build a confident understanding of the environment.
In addition to using multiple sensors, self-driving systems employ sensor fusion, which is the process of combining data from different sensors to achieve more accurate and consistent information. Sensor fusion algorithms reconcile data discrepancies from sensors and compensate for individual sensor limitations. This reduces the chances of accidents from undetected objects. Advanced neural networks are being developed to further improve sensor fusion capabilities over time via machine learning. Strong sensor coverage and fusion are vital to safely navigating complex road situations and avoiding collisions.
Once perceptions are obtained from sensors, the self-driving software (the “brain” of the vehicle) must make intelligent decisions quickly. This decision making component is another focus for safety. Researchers are developing models with built-in conservatism that prioritize avoiding risks over optimal route planning. obstacle avoidance maneuvers are chosen only after extensive validation testing shows they will minimize harm. The software also continuously monitors itself and runs simulations to ensure it is still operating as intended, with safeties that can stop the vehicle if any issues are suspected. Over-the-air updates further enhance safety as new situations are learned.
To account for any possible software or hardware faults that could lead to hazards, self-driving cars employ an entirely redundant autonomous driving software stack which is completely independent from the primary stack. This ensures that even a full failure in one stack would not cause loss of vehicle control. The redundant stack will be able to brake or change lanes if needed. There is always a fully functional human-operable primary driving mode available to fall back on. Drivers can also be remotely monitored and vehicles can be remotely stopped if any serious issues are detected during operation.
Self-driving cars are also designed with security in mind. Vehicle networks and software are tested to robustly resist hacking attempts and malicious code. Regular security updates further strengthen the systems over time. Driving data is also carefully managed to protect passenger privacy while still enabling ongoing learning and improvement of the technology. Strong cybersecurity is a fundamental part of ensuring safe adoption of autonomous vehicles on public roads.
Perhaps most significantly, self-driving companies extensively test vehicles under diverse conditions before deployment using simulation and millions of real-world miles. This gradual approach to introduction allows them to identify and address issues well before the public uses the technology. The testing process involves not just logging miles, but also performing edge case simulations, software and hardware-in-the-loop testing, redundant system checks and ongoing validation of operational design domain assumptions. Only once companies have achieved an exceptionally high level of safety are autonomous vehicles operated without a human safety driver behind the wheel or on public roads. Testing is core to the safety-first approach taken by researchers.
Through this multifaceted approach with redundant sensors and software, ongoing validation, security safeguards and meticulous testing prior to deployment, researchers are working to ensure self-driving cars can operate safely on public roads and avoid accidents even under complex conditions involving environmental changes, anomalies and unpredictable situations. While continued progress is still needed, the safety measures now in place have already brought autonomous vehicles much closer to matching and exceeding human levels of safety – paving the way for eventually preventing many of the tens of thousands of traffic fatalities caused by human mistakes each year. With appropriate oversight and care for safety remaining the top priority, self-driving cars have great potential to save lives.