Wayve Technologies

I am a research engineer working to make self-driving cars. Our approach is to take cars as actors with no script - can a vehicle drive itself in a new environment? With strange people, lighting, road signs, and the works? The challenge of building generalizable intelligence is really interesting to me - I feel like deep learning has brought so many things into the feasibility zone (like imitation learning in real life, which works ok-ish) but there are still so many things on the boundary which we care about but don't yet know how to solve (can talk for days). It's also freaky cool (and also a bit scary) to have code run on a vehicle on the actual street.

• Lots of computer vision involved. It's super important for the car not to miss anything important. And developing the tools to understand the machine learning models so that it can do this is crucial for trust. I'm quite keen on exploring more how one can create trustworthy machine learning systems.
• How does it learn behaviors? Many interesting things here regarding skill acquisition (especially if the rules change!) One of the challenges we had early on was that there was no signal for the vehicle to stay on one side of the lane! More generally, what kind of things do we explicitly care about, and what kind of things do we leave for the model to learn? If ML research has taught us anything, it's 1) priors go a long way, but 2) data is surprisingly effective.
• I need to talk to people about imitation learning and the limits of behavioral cloning. Part of my goal is to build general robotic systems which can learn new skills. (keep wishing ... ;)

NeuRender

Imagine video games, but where the graphics were rendered by a deep renderer, a neural network-powered system which would create rich, interactive renderings from thin air. Instead of creating assets for people, buildings, landscapes, etc, everything can be rendered on demand. You could potentially create virtual worlds from your own artwork. Or have a personalised experience unique to you.

Currently I'm focusing on the graphics part, specifically the part that will generate 3D worlds from 2D data. It doesn't actually need to be fully 3D; all the renderer needs to do is render convincing changes in views, and the fact that having a 3D geometric prior helps with that is a coincidence.

I'd love collaborators and/or funding. I have a Notion doc which I'll share once I separate the public and private parts, but it's roughly organised like so:

• Generative image inpainting
• Novel View Synthesis
• 2.5D and 3D related estimation

I might be wrong, but it does seem like we can rotate this image around the cloud with generative models. After all, look at StyleGAN!