The Key to Truly Humanlike Driving

The world is always changing, and that is especially true of the roads we drive on. Think of changing traffic patterns, construction, obstacles, and pedestrians. Generally, human drivers are able to adapt to the variable conditions around them — for example, to determine whether they can take a certain path around an obstacle or should wait for that obstacle to move. To put it simply, humans have the ability to manage unexpected road conditions by reasoning and by tapping into prior experiences mapped into their neural networks.

As the industry heads toward higher levels of automated driving, the challenge is to give automated systems capabilities similar to those of humans so they can make better-informed decisions and drive like humans do. That means the system has to include a reasoning system that understands and contextualizes the scene around the vehicle and can predict how that scene is likely to change in the near future.

The Right Approach to Path Planning and Scene Prediction

There are many different approaches to enabling autonomous driving. One is to use rules and heuristics, or rules of thumb: If certain conditions exist, the vehicle performs specific actions. Another is to use a detailed model of the physical world and create a system that uses encoded knowledge to mimic the choices an “expert” would make.

Aptiv’s approach is to use a state-of-the-art neural network running on an efficient artificial intelligence (AI) accelerator. Instead of using heuristics, we use extensive driving data to teach the network to contextualize its surroundings, make sense of the world, make short-term predictions about what will happen in that world and decide on its next move.

That means that we can teach it to predict the behavior of other road users. If a pedestrian is standing at the side of the road, for instance, human drivers can usually predict whether that person is about to enter the road by observing their body or head orientation. An AI/machine learning (ML) system can learn how to do that as well.

In another example, suppose the car ahead of yours has stopped in the road. Is it stopped because of a traffic light? Should you wait? If so, for how long? Should you go around it? These are questions that the system considers based on what it knows about the behavior of other road users and the environmental context around the vehicle.

The Machine Learning Advantage

Aptiv’s ML Behavior Planner system takes inputs from perception systems, mapping and localization functions, and ego-motion (the movement of the vehicle it is in). Sensors could include any combination of radars, cameras or even lidars. If a map is not available, the ML Behavior Planner can still use those other inputs to create a world model of the scene around the vehicle and determine the best path through it. Its AI/ML engine can infer lane boundaries — even if there are no lane markers — and biasing on irregular roads that lack clearly defined boundaries. In short, the ML Behavior Planner is a humanlike driver that can operate in both structured and unstructured environments.

The ML Behavior Planner outputs multiple potential driving paths that converge into a trajectory based on the overall navigational goal, safety considerations and laws. The final trajectory determines actions the vehicle will take, such as turning, changing lanes, braking or accelerating.

In this way, the ML Behavior Planner provides vehicle intelligence and behavior, regardless of the sensors and motion controllers available. Its abstraction gives OEMs the freedom to use a variety of sensors on one end and different planning technology on the other.

The baseline ML Behavior Planner can be refined to work in various regions, such as in left-driving or right-driving countries, or in places where roads are generally narrow or wide. In a sense, its driving style can be customized to match a target location.

The flexibility of the ML Behavior Planner is essential, especially as the industry continues to build steadily toward more autonomous vehicles. With additional and well-balanced data collection, the technology’s ability to handle more complex situations increases without new software architectures having to be developed. In other words, over time, the system’s capabilities improve as they get more driving experience, just like a human driver’s skills do.