Waymo has begun equipping its test fleet with a new suite of sensors, claiming improved perception and a significantly reduced cost compared with the previous generation, enabling easier scalability of its technology.
Module and system-level testing of the fifth-generation Waymo Driver platform is underway in the San Francisco Bay Area, assessing performance in different weather conditions using adapted Jaguar I-Pace SUVs.
Generational advances were developed based on experience from 20 million miles (32,000,000km) of on-road driving, and 10 billion miles (16,000,000,000km) in simulation, the company said. Despite upgrading all of the sensor packs, Waymo claims to have halved the system cost compared with its predecessor by simplifying the design and manufacturing process.
Short-range perception is improved using a new perimeter vision system, mounted at waist height on the front, sides and rear of the vehicle. It uses four lidar units to detect obstacles, augmented by six high-dynamic range cameras with overlapping fields of vision, providing additional detail needed for identification. Waymo says this enables the system to look around obstructions, quickly identify other road users and spot gaps in traffic.
The roof-mounted sensor pack features a new 360° lidar, which offers a bird’s-eye view of other road users and can detect obstacles up to 300m (984ft) ahead. Positioned above the windshield, the updated long-range camera includes a custom-designed lens and upgraded internals to provide sharper images more than 500m (1,640ft) in the distance, with improved accuracy in poor weather conditions, the company said.
Input from the lidar and cameras are overlapped with a redesigned radar architecture, with additional signal-processing capabilities and improved outputs aimed at better resolution, range and field-of-view. Waymo said this offers better detection and tracking of objects, whether moving or stationary.
Software upgrades are also in the pipeline. Waymo has developed a new model for predicting other drivers’ behavior, called VectorNet, which is claimed to require less computing power than a convolutional neural network by simplifying the maps and sensor inputs and reducing the reliance on memory-intensive data.
Traditionally systems have used HD maps rendered as images (pixels) to identify signs, lanes and road boundaries, which requires a lot of computing power. VectorNet represents the surroundings as a series of points, which can be used to create polygons for lanes, junctions and crossings, while multipoint polylines, split into smaller fragments, represent curves.
The system then uses machine learning to understand the relationships between road layouts and object trajectories, underpinning faster, more accurate predictions of traffic and pedestrian behavior, Waymo said.
Benchmarking the system against the ResNet-18 convolutional neural network, VectorNet used 71% fewer parameters and required 80% less computation in a scene with 50 agents, while demonstrating an 18% reduction in displacement error based on a data set and HD maps from Argoverse.
The software also doesn’t necessarily require complete information about its surroundings to work out what might happen next. During training, signs and markings were blanked out at random, requiring the software to fill in the gaps, as a human driver would use past experience to make predictions if some information was obscured.
“Our brains are pretty good at instinctively recognizing lines and shapes to identify objects, so it’s interesting to see this approach employed in improving autonomous vehicle tech,” commented Yuan Zhang, associate director of KPMG’s Mobility 2030 strategy. “Any advances that help improve recognition or predictive performance, or reduce the processing or data requirements needed for these sorts of tasks, are going to be helpful stepping stones toward getting to affordable, practical and safe Level 4 autonomy.
“I don’t think AVs will ever be able to deal with the full range of on-road situations, but the comparison point here should be that they need to be significantly safer and more efficient than humans, who are also not great at dealing with unpredictable – in particular, fast – scenarios.”
Waymo said it hopes improvements in speed and accuracy will increase safety and offer a smoother ride, while making the system scalable by enabling it to adapt more quickly to new operating environments.