The overall goal of REACT is a systematic, safe and validatable approach to developing, training and use of digital reality with the goal to ensure safe and reliable acting autonomous systems – especially in critical situations. In order to reach this goal, we use methods and concepts of machine learning – especially
Deep Learning and (Deep) Reinforcement Learning (RL) – to learn lower-dimensional submodels of the real world. From these submodels we compile (semi) automatically complex, high-dimensional models in order to identify and simulate the entire range of critical situations. By means of digital reality, we virtually synthesize the otherwise missing sensor data of critical situations and train autonomous systems so that they are able to handle critical situations safe and confident. The aim of the project is to enhance the capabilities of autonomous systems. Therefore we continuously and systematically validate and align synthetic data with reality and adapt the models where necessary.
In REACT, we focus on pedestrians in traffic and the optimal training of autonomous vehicles. In particular, our goal is to build motion models that are as accurate as possible covering the whole bandwidth and variability of human movements in traffic situations. We use an intelligent, agent-based simulation along with hybrid deep learning techniques, combining different techniques of Artificial Intelligence. Based on the individual submodels, we generated as automatically as possible relevant critical situations and then generate synthetic data for training.
One important aspect is the efficient and fast execution of the learning and simulation algorithms on current hardware, resulting from the use of modern compiler techniques.
Finally, we integrate the individual modules into an open system architecture.