Listen to this episode on Apple Podcasts, YouTube, Spotify, and Substack. Transcripts are also available by clicking on the “transcripts” button in the top right corner of this post on the app/web version. You can also access them from the email version by clicking here.
In today’s episode, The Driverless Digest Editor Ben Hubbard speaks with Brandon Basso, VP of Engineering at Wayve. Brandon breaks down the role of world models in autonomous driving and how Wayve is using them to train, test, and validate its AI Driver.
He explains what problems world models can solve, where their limitations still lie when it comes to rare edge cases, and how Wayve’s latest model, GAIA-4, advances the company’s approach to simulation. The conversation dives into GAIA-4’s closed-loop capabilities, how it can simulate multiple sensor modalities, and why closed-loop testing could give Wayve an edge in virtual testing and validation.
Brandon also explains the distinction between “AV 1.0” and “AV 2.0,” including how an AI-first approach changes the role of data, development, and validation. Ben and Brandon also discuss the challenge of optimizing autonomous vehicles for passenger comfort, what it takes to validate a more generalizable AI driving system, and how Wayve sees GAIA-4 playing a central role in the future of AV development and validation.
Chapters
(00:00) Introduction to Brandon Basso
(02:20) Brandon’s day-to-day role at Wayve
(03:00) What is the Wayve AI Driver?
(04:05) What is a world model, and why is it used to teach an AI how to drive?
(05:11) What problem does the world model solve?
(08:04) How some edge cases constrain what can be tested using world models
(09:21) How far off are we from capturing all edge cases using world models?
(11:35) GAIA-4, Wayve’s latest world model and its new features
(12:06) Closed-loop vs. open-loop testing with GAIA-4
(14:30) Multimodality and how GAIA-4 simulates different sensors
(15:27) How closed-loop simulation gives an edge in virtual testing
(17:38) The difference between AV 1.0 and AV 2.0
(26:13) How difficult is it to optimize for comfort in the AV industry?
(27:02) How important data is to the AI Driver
(27:38) The challenges of developing and validating an AV 2.0 system
(30:55) The future of validation and GAIA-4’s central role
(32:18) Conclusion and final thoughts
