UK automated-driving company, Wayve, has announced the launch of PRISM-1, an innovative 4D reconstruction model designed to significantly enhance the testing and training of advanced driver assistance systems and autonomous driving technology. This is the first follow-up to Wayve’s Ghost Gym, which was unveiled back in late 2023. Ghost Gym is a neural simulator that uses NeRFs to efficiently test and validate end-to-end AI driving models in a safe, controlled virtual environment.
PRISM-1 represents a significant advance in 4D scene reconstruction. It enables the resimulation of complex and dynamic scenes with minimal engineering effort, emphasizing generalization and scalability. This approach avoids the need for explicit labels, scene graphs, and bounding boxes, ensuring efficiency even as scene complexity increases. This scalability makes PRISM-1 ideal for simulating complex urban environments. Unlike traditional methods that treat vehicles as rigid entities, PRISM-1 captures their dynamic nature, including behaviors like turning indicators and sudden braking.
Interestingly, PRISM-1 is using camera-only inputs, and therefore skirts the need for LiDAR-based units. While not confirmed yet, Wayve appears to be using Gaussian Splatting with PRISM-1, as the hallmark Gaussians are occasionally visible.
While Radiance Fields are primarily known for their ability to represent a scene in hyper-realistic detail, it’s not their only feature. PRISM-1 adequately demonstrates this by identifying and tracking changes in the appearance of scene elements over time. It can further accurately resimulate complex dynamic scenarios by distinguishing between static and dynamic elements in a self-supervised manner. This extends to cyclists, pedestrians, and vehicles, which are all notoriously difficult to simulate because of their unpredictability.
Simulation plays a critical role in the development of autonomous driving. It provides a safe, repeatable, and cost-effective way to test and refine driving models, while also allowing for simulating long-tail problems. The challenge with real-world testing is that conditions can vary significantly between tests, even on the same stretch of road.
Jamie Shotton, Chief Scientist at Wayve, stated,
“PRISM-1 bridges the gap between the real world and our simulator. By enhancing our simulation platform with accurate dynamic representations, Wayve can extensively test, validate, and fine-tune our AI models at scale. We are building Embodied AI technology that generalizes and scales. To achieve this, we continue to advance our end-to-end AI capabilities, not only in our driving models but also through enabling technologies like PRISM-1.”
In conjunction with the launch of PRISM-1, Wayve has also released the WayveScenes101 Dataset, comprising 101 diverse driving scenarios from the UK and US. This dataset includes urban, suburban, and highway scenes under various weather and lighting conditions. Wayve aims for this dataset to support the AI research community in developing more robust and accurate scene representation models for driving.
This announcement comes on the heels of their $1.05 billion Series C funding announcement in early May and aligns with the outset of CVPR, where the London based team will be displaying. For those attending, they are located at booth #1712 and have created a page with their speaking schedule.
Future work will focus on further enhancing resimulation capabilities and exploring new ways to improve the accuracy and reliability of their simulation engine.
