The Purpose of the Map
There’s a scene from The Office (American) where Michael so literally follows the instructions from his GPS that he drives into the lake while screaming ‘THE MACHINE KNOWS’. While it might seem like this was written for TV, it actually happened to a woman in Tobermory, ON and it’s easy to find other unbelievable driving mishaps from people absently following online driving instructions. While this might showcase a negative effect of us blindly following our navigation guides, when is the last time you saw someone unfold a map to find their way when they are lost? We are undoubtedly dependent on our GPS but this is likely just a phase in the evolution of how humans interact with maps. As fewer and fewer people drive, our vehicles will take over orientating and navigating us to and from our destinations.
There were 700-900 years between the first known ornamental map and Ptolemy’s development of Geography in the second century AD. Even more time passed between this and the creation of more accurate maps with the dawn of a reliable compass (think the Island of California). Back then, maps were being used to define land ownership, maritime trade commerce was increasing, and there was much uncharted territory to explore. Gerardus Mercator helped with standard orientation in the 16th century but it wouldn’t be until 2016, when the AuthaGraph World Map was released, that the world would be correctly depicted, without proportion distortions, as drawing a globe accurately in two-dimensions eluded cartographers for centuries. Today, Karen Jacobsen is likely telling us where to go as our navigation maps speak to us. She is the voice for many navigation systems and helps teach us how to get around the world. Over time we learn how to get ourselves to the places we frequent without the map, although technology is making this take longer. Maps are also found in the classroom where they are an important tool in spatial reasoning and cognitive development. Children who use maps to explore their environment develop spatial skills that will help them later in life when they need to orientate themselves in unfamiliar places or solve puzzles such as mazes or jigsaw puzzles.
The most precise maps ever developed are HD maps, widely used in autonomous driving applications (ADAS systems and robotaxis). Essentially a digital twin of the built-world, it includes all sorts of information about the road environment as well as global and local positioning. HD maps have been essential for autonomous driving development, especially when computing power was a fraction of where it is today. However, it is very expensive to create HD maps (a 2018 figure had it at $5k/km) and although they are precise, they are not always accurate since they are created offline and are never a true representation of the world in real-time. Fleets of survey vehicles must drive the more than 4 million miles of US roadways collecting data which is processed and annotated in back offices. Finished maps are deployed to the ADAS/AD vehicle where it is referenced for perception. Instead of seeing its surroundings, the ADAS/AD car uses the HD map to reference what is around it and where it is within that 3D space. When the ADAS/AD car is dependent on the map in this way there is no way to learn spatial reasoning or how to get around and it will always have to reference the map for this information. For autonomous driving to become ubiquitous we must map every street – that’s a lot of miles!
One starts to question the scalability of the HD map if the process to create and maintain the map is so expansive and extensive. While some companies are focused on streamlining the map updates and maintenance through crowdsourcing, the creation of the base map remains a problem, and in time we will have mapped the entire world hundreds of times over. However, once we have mapped the world hundreds of times over, we must continue to map the world because the system was never able to learn about its surrounding 3D space. What if the system could learn from these maps so that over time it was able to interpret the inputs from the sensors and give meaning to the vehicle surroundings? What if the vehicle could get around safely without a pre-made basemap?
First, if we take away the HD map, what exactly do we need to replace? An HD map provides an autonomous vehicle with detailed information about the road network and environment. This can include information such as the location and shape of road lanes, the location of traffic signals, and the location of crosswalks and how it is supposed to interact with certain features. HD maps can also include information about the terrain, such as the elevation and slope of the road, as well as the location of buildings, trees, and other static obstacles. This information can help the vehicle to more accurately and safely navigate its environment and to plan more efficient and comfortable driving maneuvers. Additionally, HD maps can be used to improve the accuracy of the vehicle’s localization, which is the process of determining the vehicle’s position and orientation within the map.
A combination of standard GNSS + SLAM can provide both local and global localization of the vehicle. Perception models can provide object identification, location, and tracking information. How objects relate to each other and how the vehicle is supposed to respond to certain signals are part of heuristics. Heuristics can be defined as a set of rules or instructions that help the agent make decisions when faced with uncertainty or incomplete information. For example, if the agent encounters a situation where the sensor data is ambiguous, the heuristics can be used to define how the agent should act in that situation. Heuristics can also be learned by an autonomous vehicle through supervised learning and reinforcement learning.
In supervised learning, a model is trained on a labeled dataset of examples, where each example consists of input data and a corresponding label or output. For example, a model could be trained on images of road scenes and corresponding labels indicating the location of the road lanes. The model is then used to make predictions on new, unlabeled data.
In reinforcement learning the vehicle learns by interacting with its environment and receiving feedback in the form of rewards or penalties. The agent’s goal is to learn a policy, or a mapping from states of the environment to actions, that maximizes the expected cumulative reward over time.
Heuristics can also be learned by the vehicle through trial and error. For example, the vehicle can be trained in a simulated environment with a large number of different scenarios, and the agent will learn how to act based on the rewards or penalties it receives in those scenarios.
In conclusion, by using various techniques we are able to replace the functions of the HD map with real-time mapping functionality through the use of machine learning algorithms. An autonomous vehicle can drive without an HD map by using a combination of sensors, such as cameras, lidar, and radar, to perceive and understand its environment in real-time. These sensors can be used to detect and track other vehicles, pedestrians, and obstacles, and to determine the vehicle’s position and location using SLAM. The vehicle can then use this information to plan and execute safe and efficient driving maneuvers.