While the potential of autonomous vehicle technology is significant, critical shortcomings need to be addressed before its possibilities can become reality. The truth is, despite vehicles being tested across the world, autonomous vehicle technology in its current form isn’t scalable. To advance the practicality of a driverless vehicle, companies need to transition from relying heavily on HD (high definition) maps to adopting a mapless solution.
Even though nearly all significant players in the automotive industry are testing some type of autonomous vehicle, there hasn’t been much innovation when it comes to navigation. Using HD mapping to navigate autonomous vehicles is a sound approach, but the actual process of mapping an area is extremely time-consuming and tedious. Companies spend significant resources mapping certain areas, delaying deployment in difficult geographies.
What’s more challenging is the need for HD maps to be continually updated to account for changes in road structure. This can include blocked roads, construction sites, semi-static objects in the road, and more. Because of the constant mapping and remapping issue, we haven’t seen an actual commercial deployment of a driverless vehicle yet. SAE International grades autonomous driving in five levels of automation. An SAE Level 5 vehicle can drive anywhere, rendering the approach of HD mapping impractical to achieve.
Those issues can be solved by adopting an alternative navigation system that does not require HD GPS or maps. This type of mapless system requires human-grade computer vision for perception tasks like object detection, and it is also needed for path finding and “superhuman” decision-making capabilities.
Mapless technology is essentially generating accurate realizations of the car’s surroundings in real-time through machine learning, which enables vehicles to quickly and effortlessly adapt to new environments. In other words, using machine learning to generate the environment converts the challenge from a mapping challenge to a perception challenge. This architecture can be taught to generalize driving to the point where one day, an autonomous vehicle will be able to transfer its knowledge of driving in one city smoothly to another.
At the heart of the mapless solution is a learning pipeline that facilitates the understanding of the road and its obstacles in a practical manner for driving. It uses a specific architecture of neural networks that employ deep-learning methods that produce reliable vision in real-time without draining computational resources. Various sensors are mounted on the vehicle to feed the perception engine a snapshot of the world it can “see.”
The goal of a mapless self-driving car is to interpret the world logically without memorizing a specific environment. This is achieved by vetting the information the engine is trained on and generating an endless supply of artificially produced information as training data, as well as using a specific architecture of deep learning that minimizes “overfitting.” Overfitting occurs when an algorithm defined specifically for the training data can’t fit new data in to any classification, causing inaccuracies. With a developed set of tools and a sophisticated neural network, vehicles can handle the challenge of driving autonomously in a variety of environments.
A future where autonomous vehicles are the norm is possible, but not without a critical upgrade to a mapless navigation system. By transitioning from HD maps to mapless technology, we’ll start to see the full potential of these vehicles come to life. For example, by reducing the number of cars on the road and inside parking lots, we’ll be able to use the extra space for mar more important activities. The impact of driverless cars goes beyond just getting from point A to B—it will affect how humanity lives inside dense urban environments as we move toward the future.