Master Thesis: “Simulation for AI in Automotive scenarios: Interface for training and evaluation of deep learning models with simulated sensor data”
Navigation systems of autonomous vehicles rely on Artificial Intelligence (AI) algorithms to identify important traffic participants and further (static) objects of the environment. Current Deep Learning models already provide valuable approaches for the identification of lanes, vehicles, pedestrians and traffic signs. Some of the more impressive developments have been done by NVIDIA, one example is the OpenRoadNet showed below:
Detection of drivable free space around objects (Source: OpenRoadNet by NVIDIA)
However, due to the complexity of real-world scenarios, thoroughly testing such AI-based perception systems as part of the navigation systems of autonomous vehicles can be very costly. This is where using Hybrid Testbeds (HTBs) can become valuable. These HTBs fuse simulated and real-world components to enable system tests within safe and reproducible virtual test drives and thus complement and enhance current testing pipelines. In order to use HTBs for this purpose, a framework that allows the training and evaluation of embedded AI-Platforms needs to be developed. For this, the following goals are to be fulfilled:
- Literature research with focus on X-in-the-loop, AI training and inference.
- Conceptualize an interface for training and inference of Deep Learning frameworks using standard practices (e.g. ASAM Open Simulation Interface standard) and metrics.
- Development and implementation of the interface as a plugin in the 3D simulation software VEROSIM.
- Application and evaluation of the framework (e.g. by using the NVIDIA DriveNet and OpenRoadNet DL models)
Desirable goals would also be:
- Implementation and evaluation of the framework in hardware setups, using the NVIDIA Xavier AGX.
Keywords: Autonomous driving, Deep Learning training, X-in-the-loop
Supervisor: Jiménez