We’re using traffic cameras, accident reports and insurance claims from around the world, along with stories submitted by the public, to compile the world’s largest library of edge case scenarios.

To date, AVs have mostly been trained to spot entire vehicles and pedestrians in well-lit and predictable environments. But the elephant in the room here could quite literally be the elephant on the road: how will an AV respond to unusual occurrences? While AV developers have trained their systems on a huge number of animal-related scenarios, there’s a nearly infinite number of ways that the real world can trip up AVs.

Training and retraining AVs using edge case simulations means testing them to their limits.

It involves teaching them to recognise high-risk events (a construction worker emerging from a manhole cover, pedestrians in dappled light, objects falling from a motorway bridge), and the predictors of potentially hazardous events too – like cars peeking out from behind other vehicles, pedestrians stepping out from behind buses in traffic, or cyclists running red lights concealed by other vehicles.

The library is used to train and retrain autonomous vehicles (AVs), as well as to develop examples of edge cases that haven’t occurred or been recorded yet.

Edge cases are organised into a knowledge graph, which provides a clearer understanding of the entire risk space. With this information, it’s even possible to extrapolate to other scenarios that haven’t been discovered or recorded yet – and then to use these to train and retrain until the AV can demonstrate the safest possible response.