Section: Application Domains

Autonomous Vehicle

This new application domain builds in fact upon former collaborations of the TAO team with the automotive industry, that created the links with some of the researchers of the R&D departments of Renault (within the Systematic CSDL project and the SystemX ROM project (François Gonard's PhD) and PSA (M. Yagoubi's PhD [84], [85]).

The current work, in collaboration with Renault, is related to the safety of the autonomous vehicle. The validation of the software system is today based on statistics of incidents (failures of some automatized component) assessed from millions of hours of 'driving', either by human drivers in real cars, or by simulations. The work for TAU is related to the set of sample scenarii that are used to compute these statistics. This will require in the first place to identify some latent representation space common to both the actual real-life experiments and the results of the simulation, something that will be achieved using Deep Auto-Encoders of the time series recording the experiments. Two works have started this Fall:

• How to assess the representativity of current set of scenarii, and identify new scenarii to be fed into the simulator to improve the coverage of the scenario space in the common latent representation space, and is the goal of the yet-to-be-signed POC with Renault (Raphaël Jaiswal is working on Renault data since September 2017);

• How to identify original scenarii that lead to failures, an optimization problem in the scenario space. Several criteria for failures will be considered (e.g., getting too close to the preceding car), and the optimization will most likely require building a surrogate model of the simulator for each chosen criterion (and here again Deep Networks are a good candidate), due to its high computing time. This is the topic of Marc Nabhan's CIFRE PhD, started in October 2017 (after a 3 months internship).