Today, a week after testing on public roads resumed and days after $ 750 million in capital was raised, Waymo has abolished an AI model that it claims is the ability of its driverless systems to control pedestrian behavior, Cyclists, predict, has “significantly” improved. and driver. Called VectorNet, it supposedly delivers more accurate projections and requires less computation compared to previous approaches.
Anticipating the future positions of road agents is a table for driverless cars that by definition have to navigate challenging environments without human supervision. As the collision with an autonomous Uber vehicle and a cyclist in March 2018 tragically shows, perception is crucial. Without them, self-driving cars cannot reliably decide how to react in known or unknown scenarios.
VectorNet is designed to help predict the movement of road users by creating representations to encode information from maps, including real-time trajectories. Like the rivals Cruise and Aurora, Waymo collects high-resolution, centimeter-precise maps of regions in which its autonomous vehicles drive. Together with sensor data, these provide a context for the Waymo driver, Waymo̵
This is where VectorNet comes in. In contrast to the replaced convolutional neural networks, which were operated with computing-intensive pixel renderings of cards, VectorNet records each card and each sensor input in the form of vectors (sketches of points, lines and curves based on equations based on mathematical data).
Waymo uses vectors to represent road features as points, polygons, and curves. Lane boundaries contain multiple points that form a spline (i.e. curves that add up to larger continuous curves), crosswalks are polygons with at least two points, and stop signs are represented by a single point. These geographic units can be approximated by polylines (connected series of line segments) consisting of points together with their attributes, while moving agents can be estimated by polylines based on their trajectories.
Graphic neural networks work directly with graphs or mathematical objects that consist of nodes and edges. In VectorNet, a neural network with a hierarchical graph, each vector is treated as a node, and data from the maps is passed along with the agents’ trajectories to a target node across the network. An output node corresponding to the target agent is used to decode the trajectories.
VectorNet first receives information at the polyline level before it is passed on to a diagram to model higher-order interactions between the polylines. It calculates the future trajectories of objects and records the relationships between vectors, e.g. For example, when a car enters an intersection or when a pedestrian approaches a zebra crossing, which can help predict the behavior of agents.
To further improve VectorNet’s capabilities and understanding of the world and thereby improve its predictions, Waymo trained the system to learn from contextual information and draw conclusions about it could happen near a vehicle. The company’s researchers accidentally hid map features during the training, e.g. For example, a stop sign at an intersection in four directions, and VectorNet asked to complete the missing items. In validation tests using Waymo’s own data set and Argo AI’s Argoverse, VectorNet performed 18% better than ResNet-18 (a popular convolution network), using 29% of the parameters (variables) and consuming 20% of the calculation on average.
“These improvements allow us to make better predictions and offer our drivers a safer and smoother experience, and even packages that we transport on behalf of our local delivery partners,” said Waymo in a statement. “This will be particularly beneficial as we expand into more cities, where we will continue to encounter new scenarios and behaviors. With VectorNet, we can better adapt to these new areas, learn more efficiently and effectively, and achieve our goal of making fully self-driving technology available to more people in more places. “
This is not the first time Waymo has used AI to accelerate workloads like perception, data expansion, and search.
In early April, the company announced Progressive Population Based Augmentation (PPBA), a system that it claims has improved the performance of its object recognition systems while reducing the amount of data needed to train. Waymo worked with DeepMind on PBT (Population Based Training), which has managed to reduce false alarms in detection tasks for pedestrians, cyclists and motorcyclists by 24% while at the same time halving training time and computing resources. And Waymo previously put content search in the spotlight, based on technologies similar to Google Photos and Google Image Search, so data scientists can quickly locate almost every object in Waymo’s driving history and logs.