Wayve has launched PRISM-1, its new 4D reconstruction model designed to enhance the testing and training of its ADAS and autonomous driving technologies. Through the model, Waze is looking to develop scalable, realistic, re-simulations of complex driving scenes with minimal engineering or labelling input.

Initially showcased in December 2023 through its Ghost Gym neural simulator, Wayve has used novel view synthesis to create precise 4D scene reconstructions using only camera inputs and a method that it says will revolutionize autonomous driving simulations by accurately and efficiently simulating the dynamics of complex, unstructured, real-world environments. PRISM-1 utilizes novel view synthesis techniques to accurately depict moving elements, such as pedestrians, cyclists, vehicles and traffic lights, and precise details like clothing patterns, brake lights and windshield wipers.

While powering the next generation of Ghost Gym simulations, PRISM-1 also uses a flexible framework that helps identify and track changes in the appearance of scene elements over time. This enables it to more accurately re-simulate complex dynamic scenarios with elements that change in shape and move throughout the scene. Here, it can distinguish between static and dynamic elements in a self-supervised manner, reducing the need to implement explicit labels, scene graphs and bounding boxes to define the configuration of a busy street, for example. This approach maintains efficiency, even as scene complexity increases, ensuring that more complex scenarios do not require further engineering effort. Wayve says that this approach makes PRISM-1 a scalable and efficient solution for simulating complex urban environments.

In addition to launching PRISM-1, Wayve also rolled out its WayveScenes101 Benchmark dataset. The dataset is comprised of 101 diverse driving scenarios from the UK and US, including urban, suburban, and highway scenes under various weather and lighting conditions. Through it, Wayve is looking to support the AI research community in both advancing novel view synthesis models and accelerating the development of accurate, more robust, scene representation models for driving.