This self-driving startup is using generative AI to predict traffic

Autonomous driving technology has traditionally relied on machine learning for route planning and object detection. However, a new wave of companies and researchers are turning to generative AI to advance autonomy further. Wayve, a competitor of Waabi, introduced a model trained on video data collected by its vehicles last year, marking a significant development in the field.

Waabi’s approach involves utilizing generative AI technology similar to models like OpenAI’s DALL-E and Sora. Their model, Copilot4D, analyzes point clouds of lidar data to create a 3D map of the car’s environment, breaking it down into sections for analysis, akin to how image generators process images into pixels. By continuously predicting the movement of all points of lidar data, Waabi’s system can anticipate scenarios 5-10 seconds into the future.

The company, along with competitors Wayve and Ghost, adopts an AI First approach in developing autonomous driving systems. This methodology involves systems that learn from data rather than being pre-programmed for specific reactions to situations. By focusing on this data-driven approach, these companies aim to potentially reduce the need for extensive road-testing of self-driving cars, a hot topic following an incident involving a Cruise robotaxi in October 2023.

Waabi distinguishes itself from rivals by focusing on generative models tailored for lidar data rather than cameras. According to Urtasun, a key figure in the company, lidar technology is crucial for achieving Level 4 autonomy, where the vehicle can operate safely without human intervention. Cameras excel at visual representation but fall short in accurately gauging distances and comprehending the geometry of the car’s surroundings, making lidar technology indispensable for a comprehensive autonomous driving system.