Waabi, along with competitors Wayve and Ghost, are adopting an "AI-first" approach, focusing on systems that learn from data rather than being programmed for specific situations. Unlike its competitors, Waabi is building a generative model for lidar, not cameras, as CEO Raquel Urtasun believes lidar is essential for Level 4 automation. However, the company will not use the model's generated videos as training in its driving simulator to avoid teaching any "hallucinations" from Copilot4D. The technology is not new, but it's the first time a generative lidar model has been scaled up for commercial use.
Key takeaways:
- Self-driving company Waabi has developed a generative AI model, Copilot4D, that predicts the movement of vehicles based on data from lidar sensors.
- The model breaks down lidar data into chunks and predicts how all points will move, generating predictions 5-10 seconds into the future.
- Waabi's approach is described as "AI-first", meaning the system learns from data rather than being taught reactions to specific situations. This could potentially reduce the need for extensive road-testing of self-driving cars.
- While the technology behind Copilot4D is not new, it is the first time a generative lidar model has been scaled up for commercial use. However, Waabi has not yet decided whether to make the model open-source.