‘DeepSafe’ a dRISK led consortium have been awarded £2m to unlock next phase in the evolution of safe, autonomous vehicles
5th Sep 2023 dRISK
A team of leading academics and researchers in AI, simulation, strategy and engagement firms has been awarded £2 million in CCAV funding to advance DeepSafe, the next stage in the commercialisation and deployment of self-driving vehicles.
The dRISK.ai-led consortium of DG Cities, Imperial College London, Claytex and rFpro will unlock a barrier in the supply chain – together, they will develop the simulation-based training needed to train autonomous vehicles (AVs) to handle ‘edge cases’, the rare, unexpected driving scenarios they must be prepared to encounter on the road.
DeepSafe will commercialise ‘sensor real’ edge case data – a simulation of what an actual sensor would detect – together with AV training tools, for release in the UK and internationally after the project. As well as advancing self-driving systems, the grant aims to support innovation in industry, job creation, and investment, building the capacity to develop AV technology in the UK and export it to the rest of the world.
Rav Babbra, dRISK.ai: “There are currently bottlenecks in the technology sector that are inhibiting deep-learning using simulation. These are what the DeepSafe consortium aims to resolve. Together, we will enable the performance improvements in AI perception needed to develop safer and more reliable autonomous driving behaviours.”
Pilot studies have demonstrated that AVs can handle the straightforward situations that make up 99% of everyday driving experiences. However, it is more difficult to train vehicles to deal with ‘edge cases’, the rare and unusual events that can happen on the road. Failure to understand an edge case can result in unreliable, unsafe behaviour, such as phantom braking, lane-keeping failure and collisions.
To train AVs to safely navigate these situations, they have to be identified and simulated. Part of the team’s previous project, DRISK focused on engagement, using the public’s driving experiences to crowd source these rare events and build scenarios – a process that also served to raise awareness of the technology and give an insight into consumer barriers to autonomy.
However, such real-life training data is limited. There is a widely recognised need for simulation and synthetic data, but the artificial data currently available is not sufficiently ‘sensor-real’ for AI trained by simulation to improve its perception and decision-making. DeepSafe brings together leaders in the simulation supply chain to resolve these synthetic data issues to enable the successful simulation-based training vital to develop safe, reliable self-driving services.
In addition, DeepSafe will establish the definitive toolchain to realistically represent the dynamics of vulnerable road users, and will conclusively answer ‘how close does simulation have to be?’ to train an ADS (Automated Driving System) to outperform a human driver by an order of magnitude.
Complementing the new data and simulation techniques will be a novel “AI Canvas” – a new kind of software for understanding the weak points of learned systems, backtracking failure modes to their source, and correcting them with new training data. With this complete toolchain, the project hopes to finally unblock the way to safe and economical AV deployments.
Balazs Csuvar, Head of Delivery at DG Cities: “A lot of our focus within the consortium is on user perception – how can we make sure that people feel safe in self-driving cars, how can they be reassured by and trust the work that has gone into the testing. Lane-keeping assist and other ADAS features (in some way precursors to fully autonomous driving) already influence the way we drive, without people actually knowing how well they perform. We will work on understanding how best to communicate to drivers this crucial information and use it to outline how autonomous systems should be benchmarked as well.”