Back to ASAM International Conference 2019

Extraction of real-world driving scenarios and generation of effective-valid scenarios for Autonomous vehicles and ADAS using artificial intelligence 

Mr. Harsha Jakkanahalli Vishnukumar  
Project Leader and Architect AI-Core
AKKA GmbH & Co. KGaA  

Autonomous driving for vehicles is one of the most important and revolutionary technology. Autonomous driving is expected to increase road safety, improve mobility, and will establish new opportunities and branches in the industry. With the advancements and complexity of such driver assistance systems increases, the efforts required for testing, Verification and Validation also increases. Each function or feature in the vehicle has to work without failure and finally translate into no accidents until the end of life of the vehicle, which means billions of kilometers to be test driven in the real-world successfully before validating and final release of autonomous vehicles. In addition to which we have incredible number of eventualities and unpredictable situations in the real-world. Since each direct or indirect traffic participant has a large set of possible motions and possible combinations of these motions, which can lead into unexplored areas of situations. This could possibly occur in the real-world, leading to different interactions with the autonomous vehicle, which has not been encountered yet. Hence we propose an effective method with AI-Core (Artificial Intelligence Core) to extract real-world scenarios from real-world reference sensor data scene-by-scene, and additionally extract high-level feature vectors in the latent space to explore and generate valid dynamic length scenarios using Variational-Autoencoder-decoder networks. Additionally scenarios concentrating on bugs and failures can be generated within the simulated environment using reinforcement learning, covering possible valid test cases which can be seamlessly using in both laboratory XIL tests and real-world automated tests. This paper will include examples and experiments conducted to extract scenarios from real-world sensor data and experiments to generates valid scenarios, which are later induced into a Simulation environment for Verification.