Season 2 Ep. 8 Alex Kendall of Wayve on teaching cars to drive with machine learning

Summary

On the Season 2, Episode 8 of the Dr. Pawd podcast, Alex Kendall, co-founder and CEO of Wayve, a London-based company pioneering AI technology for autonomous driving, shared his insights on teaching cars to drive with machine learning. Wayve is building AV 2.0, a next-generation autonomous driving system that can quickly and safely adapt to new driving domains anywhere in the world. The company has announced commercial pilots with two of the UK's largest online grocery retailers and is backed by a $200 million series B round from Microsoft and Virgin.

Alex Kendall's fascination with being able to teach robots to see in real-time and make advanced decisions led him to the idea of building autonomous vehicles. Wayve's team and idea started in a garage in Cambridge with computer equipment and a car that could learn live on the road with just a few examples. The company believes they can build something that can behave in more intelligent ways than traditional autonomous vehicles and coexist with other human drivers in a natural way.

The traditional rule-based systems cannot effectively bridge the gap to achieve this level of intelligence. Machine learning is the future of self-driving. The end-to-end approach entails replacing this architecture with a single giant neural network that can process sensory input and output motion plans that are scalable to different driving domains and vehicles. The approach requires a considerable amount of training data and infrastructure to engineer the model and train the system.

The system can classify at every state where in that high-dimensional matrix that point is, which gives them the ability to understand performance at a very granular level and compare like-for-like, factor out the difficulty, and compare performance. Safety drivers serve as driving instructors for the robots and help find interventions and learning examples. They measure performance within a target deployment domain and look at how successfully the system can drive autonomously in a safe and trustworthy way.

The sensing platform used should be the most safe and scalable as the technology can learn to drive based on data. Visual spectrum is the primary input for autonomous vehicles, but other sensors like inertial, GPS, and microphones are important too for decision-making. Advances in deep learning, such as the ability to quantify the uncertainty of a neural network and understand where the attention lies or where the saliency of a decision is, now give them the ability to build a system they can really understand.

The two important things necessary for an intelligent system on the road are the ability to understand uncertainty and quantify what the system knows and doesn't know, and looking at behaviors on a collective level rather than hand coding rules for specific events. Contextual understanding and awareness of the full scene is necessary to make decisions on the road. Perception and control cannot be separated out since an interface that's hard coded can be limiting. Instead, control needs access to everything necessary to make decisions, and this can be achieved by letting the interface be learned.

The most impactful and interesting use case for autonomy is operating in urban domains, where most of the populations are centered and where energy and transportation exist. For this reason, tackling the hardest problem first, i.e., operating in urban domains, forces us to think very deeply about the problem and build a solution that will impact and provide value to the world. Autonomy is a pivotal problem for society as it will be the first time machines and artificial intelligence will have a physical interface with society.

Embodied autonomy, embodied intelligence, and autonomous mobile robotics will be the next wave of computing that will be as transformative as the personal computer or the iPhone was. This new paradigm where we have these mobile robots that move around about the world, connected and with sensing on board, will allow us to move people and goods around more effortlessly, sustainably, and safely, and will provide us with a platform to build levels of intelligence that can really allow us to uncover and understand more about what we do in society.