We combine state-of-the-art computer vision with Imitation Learning to build realistic models of human behaviour. By extracting the “latent logic” behind human actions, our models respond realistically even in situations they have not seen before. Our software is platform agnostic and runs on lightweight compute resources, enabling scalable testing.


Computer Vision

Our behaviour models learn from demonstrations that are collected from real-life examples of natural human behaviour.

We collect these by applying our cutting-edge Computer Vision, which we developed in-house and exceeds the current published state-of-the-art, to video data from traffic cameras, drone footage, and soon, from on-car cameras. Our Computer Vision detects and classifies road users, tracks their motion, and maps it from the pixel-space of the video into a position in the real three-dimensional world.

Our models learn to generate realistic trajectories from this real-life behaviour. By training our models on data extracted from certain types of road users or traffic contexts, we can create type- and situation-specific behaviours.

Real Data
No Manual Labelling

Imitation Learning

Our models are realistic because, rather than hard-coding a specific set of behaviours, we apply machine learning to create agents that develop their own behaviours based on how humans actually behave. Because an a-priori description of human behaviour in all possible situations is almost impossible, agents with hard-coded behaviour cannot have the same flexibility and adaptability as our AI agents.

Unlike traditional machine learning, such as reinforcement learning, which uses a hand-crafted “reward signal” to train the machine, our models learn by imitating the behaviour of an expert. Removing a manually programmed reward from the learning process allows the models to learn natural behaviour, rather than finding behaviours that might be better at maximising their particular reward but are nonetheless different from how a human would behave. Using Imitation Learning lets us create realistic rather than perfect behaviour models.

No Hard-Coding
Easily Generated
Realistic Models

Watch an overview


Scenario Generation

Imitation Learning provides the flexibility to learn behaviours specific to a certain context, incorporating the subtle variations in behaviour that humans show in different situations.
This includes behaviours for different road user types, road layouts and locations. We’re working on building behaviours for different driving styles and driver types.
As well as running tests under “normal” road conditions, we also support edge-case scenario generation.

Edge Cases
Driving Styles

Simulation Integration

Our agent models are designed to be lightweight and run on the cloud, enabling millions of miles of simulation overnight. Hundreds of agents can be modelled simultaneously on a standard laptop.

Our agent model can be implemented in any commercially available or proprietary platforms, giving our customers the ability to continue to work with their known and trusted tools, while gaining new functionality.

Integrating with our software is easy – it’s based on the standard API used by existing agent model interfaces and can work on any simulation platform. We offer 2 integration approaches: API-based and co-simulation.

Platform Agnostic
Standard API
Lightweight Compute