We’re solving some of the most complex and exciting problems in end-to-end autonomous driving research.

Diverse autonomous driving scenes

Driving the science of autonomy

Since the very beginning, Wayve has been based on cutting-edge research advances in artificial intelligence and computer vision. Today our applied research division, Science remains at the forefront of end-to-end autonomous driving research.

Working across several subfields of AI and learned decision-making, we push the boundaries of what’s possible with machine learning to give step changes in how our self-driving cars perform on the road—advancing the state of the art in embodied intelligence.

World modelling

We are developing algorithms that allow our robotic cars to understand the world around them. Generative models are one promising approach for reaching this goal.

Training generative models involve collecting a large amount of real-world driving data and then using latent models to understand the complex dynamics of what’s happening in the scene to generate new data like it. From these ‘world models’, we create techniques that have broad applications to improve motion planning for self-driving vehicles.

Our science teams are at the forefront of this research. Our paper at NeurIPS 2022 showed how our world models could improve the generalisation of driving policies, making self-driving cars drive better.

Read our blog

Download the research paper

Learned Agents

Learned agents

To build an effective simulator to train and properly test our driving software, we need virtual road users to behave the same way in simulation as they would in the real world. This requires new techniques to simulate good and bad driving, including the “long tail” of unusual, weird, and rare edge cases.

We’re developing new research around multi-agent reinforcement learning to populate the Wayve Infinity Simulator with the realistic behaviour of other vehicles, cyclists, buses and pedestrians—all acting unpredictably based on their goals and conventions.

Our team is exploring model-free and model-based reinforcement learning methods and transformer-based architectures for self-driving. They’re also working to improve training efficiency and measure how realistic and diverse their agents are.

Read our blog

Watch the video below to see how we simulated a busy intersection with a mixture of mild and aggressive learned agents controlling cars and bicycles.

Multi-modal foundational models for driving

Large language models (LLMs) and large vision-language models (VLMs) have emerged as powerful general-purpose foundational models that can be used to drive up the performance of semantic tasks across many domains.

At Wayve, we’re inspired by these developments and envision a future where new multi-modal foundation models for embodied intelligence will enable the training of driving models using vision, language, and action (VLA) supervision. We are investing in large-scale VLA models capable of both high-level reasoning as well as planning in natural language and corresponding low-level continuous control prediction.

Automatic generation of 3D worlds

Neural rendering

Neural rendering

Recent advancements in neural rendering have shown great promise in using neural networks to represent 3D scenes more efficiently and robustly. We’re using this technique to automatically generate photorealistic 3D worlds and scenarios from real-world driving data.

We can use these scenarios to create unit tests for our learned planner before we put it on the road. This is far less expensive than performing on-road testing. Furthermore, we can augment these base scenarios in many ways to create novel scenarios that can be used as training data for our learned driving policy.

simulated driving scene

Scaling simulation

We’re also exploring ways to improve the volume and diversity of data in our procedural simulator, Wayve Infinity Simulator. Our team is working to bridge the gap between simulation and reality by offering new techniques to efficiently scale our training corpus with photorealistic, agent-realistic, and world-realistic synthetic data.

This allows us to create a limitless number of diverse worlds for the Wayve AI Driver to experience, enabling us to robustly test our driving models across challenging or safety-critical edge cases.

Science Leadership

Picture of Jamie Shotton
Jamie Shotton
Chief Scientist

“Science is at the heart of Wayve’s ability to innovate in how we build neural networks for autonomous decision-making. It’s inspiring to see how our own research ideas, built on the shoulders of continued advances in the wider research community, are giving us step changes in AI capability leading to safer, more performant, and more generalisable driving.”

Meet Jamie

Picture of Vijay Badrinarayanan
Vijay Badrinarayanan
VP of AI

“We’ve now entered a decade where the key challenge is to build embodied intelligence, which means agents that can learn to work in real everyday environments with human activity. This brings new and exciting challenges to tackle as researchers.”

Meet Vijay

Check out more publications supporting cutting-edge end-to-end autonomous driving research.

Discover more

Get in touch

If you want to work with incredible people to solve the most exciting problems in end-to-end autonomous driving, check out our jobs.


Principal Scientist
Principal Research Engineer

Off-Board Software

Senior Platform Engineer, ML platform