Wayve co-founder and CEO Alex Kendall sees promise in bringing his autonomous vehicle startup’s tech to market. That is, if Wayve sticks to its strategy of ensuring its automated driving software is cheap to run, hardware agnostic, and can be applied to advanced driver-assistance systems, robotaxis, and even robotics.
The strategy, which Kendall laid out during Nvidia’s GTC conference, begins with an end-to-end data-driven learning approach. This means that what the system “sees” through a variety of sensors (like cameras) directly translates into how it drives (like deciding to brake or turn left). Moreover, it means the system doesn’t need to rely on HD maps or rules-based software, as earlier versions of AV tech has.
The approach has attracted investors. Wayve, which launched in 2017 and has raised more than $1.3 billion over the past two years, plans to license its self-driving software to automotive and fleet partners, such as Uber.
The company hasn’t yet announced any automotive partnerships, but a spokesperson told TechCrunch that Wayve is in “strong discussions” with multiple OEMs to integrate its software into a range of different vehicle types.
Its cheap-to-run software pitch is crucial to clinching those deals.
Kendall said OEMs putting Wayve’s advanced driver-assistance system (ADAS) into new production vehicles don’t need to invest anything into additional hardware because the technology can work with existing sensors, which usually consist of surround cameras and some radar.
Wayve is also “silicon-agnostic,” meaning it can run its software on whatever GPU its OEM partners already have in their vehicles, according to Kendall. However, the startup’s current development fleet does use Nvidia’s Orin system-on-a-chip.
“Entering into ADAS is really critical because it allows you to build a sustainable business, to build distribution at scale, and to get the data exposure to be able to train the system up to [Level] 4,” Kendall said onstage Wednesday.
(A Level 4 driving system means it can navigate an environment on its own — under certain conditions — without the need for a human to intervene.)
Wayve plans to commercialize its system at an ADAS level first. So, the startup designed the AI driver to work without lidar — the light detection and ranging radar that measures distance using laser light to generate a highly accurate 3D map of the world, which most companies developing Level 4 technology consider to be an essential sensor.
Wayve’s approach to autonomy is similar to Tesla’s, which is also working on an end-to-end deep learning model to power its system and continuously improve its self-driving software. As Tesla is attempting to do, Wayve hopes to leverage a widespread rollout of ADAS to collect data that will help its system reach full autonomy. (Tesla’s “Full Self-Driving” software can perform some automated driving tasks, but isn’t fully autonomous. Though the company aims to launch a robotaxi service this summer.)
One of the main differences between Wayve’s and Tesla’s approaches from a tech standpoint is that Tesla is only relying on cameras, whereas Wayve is happy to incorporate lidar to reach near-term full autonomy.
“Longer term, there’s certainly opportunity when you do build the reliability and the ability to validate a level of scale to shrink that [sensor suite] down further,” Kendall said. “It depends on the product experience you want. Do you want the car to drive faster through fog? Then maybe you want other sensors [like lidar]. But if you’re willing for the AI to understand the limitations of cameras and be defensive and conservative as a result? Our AI can learn that.”
Kendall also teased GAIA-2, Wayve’s latest generative world model tailored to autonomous driving that trains its driver on vast amounts of both real-world and synthetic data across a broad range of tasks. The model processes video, text, and other actions together, which Kendall says allows Wayve’s AI driver to be more adaptive and human-like in its driving behavior.
“What is really exciting to me is the human-like driving behavior that you see emerge,” Kendall said. “Of course, there’s no hand-coded behavior. We don’t tell the car how to behave. There’s no infrastructure or HD maps, but instead, the emergent behavior is data-driven and enables driving behavior that deals with very complex and diverse scenarios, including scenarios it may never have seen before during training.”
Wayve shares a similar philosophy to autonomous trucking startup Waabi, which is also pursuing an end-to-end learning system. Both companies have emphasized scaling data-driven AI models that can generalize across different driving environments, and both rely on generative AI simulators to test and train their technology.
