Published on March 15, 2024, 11:17 am

Waabi, a self-driving company, has revealed its utilization of a generative AI model to predict vehicle movements. The innovative system, named Copilot4D, is trained on extensive data from lidar sensors, which rely on light to gauge distances of objects. By inputting scenarios like a fast-merging driver onto a highway, the model anticipates surrounding vehicle actions and generates a future lidar representation (potentially indicating accidents) within 5 to 10 seconds. Although the current version of Copilot4D is the focus of today’s announcement, Waabi’s CEO Raquel Urtasun mentions an enhanced and more interpretable version being tested in autonomous trucks in Texas.

While traditional autonomous driving systems have leaned on machine learning for route planning and object detection, there is an emerging trend among companies and researchers towards leveraging generative AI models for advancing autonomy. Wayve, one of Waabi’s rivals, introduced a similar model last year trained on video data collected by their vehicles.

Waabi’s model operates akin to image or video generators such as OpenAI’s DALL-E and Sora but tailored for lidar data capture instead of cameras. The system dissects point clouds received from lidar data into segments based on its training to make predictions about the movement of all points continually to forecast up to 10 seconds ahead.

Differentiating itself from competitors, Waabi opts for a generative lidar model over camera-based approaches with the belief that lidar is crucial for achieving Level 4 automation where human attention during driving isn’t necessary due to its clear capability in measuring distances accurately compared to cameras.

The strategic emphasis at Waabi centers around an “AI-first” approach aiming at systems that learn directly from available datasets rather than predefined reactions to circumstances. They anticipate this method could streamline road-testing hours needed for self-driving car validation following critical incidents like the flagged Cruise robotaxi incident involving a pedestrian being dragged.

While Bernard Adam Lange acknowledges the existing technology basis behind these models from his research experience at Stanford University, he emphasizes that scaling up generative lidar models like Waabi’s into commercial applications marks significant progress towards more rapid and precise reasoning capabilities in autonomous vehicles for various decision-making tasks.

Despite its impressive predictive capabilities extending up to about 10 seconds ahead with particular focus on 3-second projections according to benchmark tests highlighted by Waabi, there are considerations about the usefulness when projecting further ahead as outlined by Chris Gerdes from Stanford’s Center for Automotive Research.

A discourse persists within the field debating whether generative AI models should be open-sourced like Copilot4D since releasing it could support academic research efforts seeking insights into safety evaluation and progressing the technology while also enabling scrutiny by competitors. Urtasun values academia’s role in shaping self-driving innovations advocating for transparency through open-source models yet balancing prudence against revealing proprietary advancements too soon.

In conclusion, Waabi’s pioneering use of generative AI in their Copilot4D system not only enhances predictive capabilities but also stimulates broader discussions around ethical considerations and collaborative advancement prospects within the realm of autonomous driving technologies.


Comments are closed.