The funding project KI Delta Learning has provided new research results on scalable AI for automated driving. On March 9 and 10, 2023 the 17 project partners presented the research results of the cooperation project during the final event at the Mercedes-Benz premises in Stuttgart- Vaihingen. Focus was on keynotes, expert talks and poster presentations. The first day targeted the public; the second day was reserved for the members of the KI Familie. More than 115 participants on site and dozens online gained an interesting insight into the more than 3-year lasting project work. Officially, the project will end by the end of April 2023.
Highly and fully automated vehicles are facing a large variety of complex situations in a continuously evolving world of mobility. Especially for environment detection (perception), Artificial Intelligence is a key technology. In recent years, AI made huge progress in this respect.
So far, automotive AI was trained for limited scenarios only, for instance such as “highway driving during good weather”. In this scenario (domain) AI is safe and reliable. To work in other environments, for instance such as “highway driving during rain”, AI algorithms needed re-training for this new domain, resulting in enormous development costs. The KI Delta Learning project therefore has also investigated new approaches in machine learning enabling a more efficient training of AI modules.
Efficient AI Training Methods: Learning Deltas
By transferring existing knowledge, the project aimed at learning deltas – meaning different requirements between a familiar domain and a new target domain. Learning deltas helps closing current gaps that limit the Technology Readiness Level (TRL) of autonomous vehicles and slow down a broad application of AI in autonomous driving.
The project investigated, developed and applied disruptive methods for AI training allowing continuous learning in a more sustainable way. Knowledge that has already been learned as well as previously tested and secured development levels are retained when changing domains. This represents an efficient approach to keep up with ever-shorter innovation cycles and the challenge of constantly changing mobility systems.
Deltas addressed in the project included changes in sensors, different traffic areas – from country roads to complex city traffic, different countries, different daytimes, seasons and weather, long-term traffic changes by new mobility concepts and road users, and last but not least ongoing development of AI methods such as better training strategies and more efficient neural networks.
Project’s Key Areas: Data, Transfer Learning, Didactics, Automotive Suitability
To provide a basis for development in the three areas, a project-specific data set tailored to the project objectives was produced and labelled. This is required to cover all deltas explored in the project and being compliant to data protection. Therefore, a real word data and a synthetic reference dataset was created.
Transfer learning is powerful machine learning technique that enables the reuse of knowledge acquired from one domain to another. This technique has been gaining importance in the automotive field, as it allows for the utilization of existing knowledge from other domains to improve the accuracy of autonomous vehicle systems.
Didactics enables learning by structuring the learning process. While in principle it is clear how to train a neural network with annotated data, the acquisition of the data is time and resource consuming, as human annotators are required. Therefore, in the project new approaches to learn with only a few data points annotated by humans or with no such labels at all were developed. This was complemented by approaches to accelerate learning by optimizing network architectures as well as optimized learning processes.
The process commonly used in the industry to engineer automotive AI systems is to develop, train, and verify AI functions in the lab using recorded data. This causes two problems: First, the high-performance computer hardware in the lab and the embedded hardware in the vehicle differ significantly, and second, the situations in which a vehicle encounters in the real world may differ significantly from the previously recorded training and test data. Both problems were addressed in the KI Delta Learning project, in order to guarantee the reliability of AI systems for automotive driving functions as well as to increase the robustness of AI systems even in the presence of unexpected or unknown scenarios.
In the KI Delta Learning project, more than 300 people teamed up and published more than 90 scientific papers.
Collaboration in Artificial Intelligence
KI Delta Learning is part of the four projects of the KI Familie (KI Absicherung, KI Delta Learning, KI Wissen und KI Data Tooling), initiated and developed by the VDA Leitinitiative Connected and Autonomous Driving. 80 leading partners from science and industry are involved receiving funding from the Federal Ministry for Economic Affairs and Climate Action (BMWK).
In this unique setting, all KI Familie projects are working together. The partners are sharing knowledge while fostering pre-competitive collaboration which is essential in an ever more competitive and complex environment with fast pace innovations. Exchanging findings across project boundaries accelerates the knowledge buildup in cutting-edge technologies for the good of industries, research institutions and society. The joint commitment to share precompetitive knowledge helps each partner to stay technologically ahead and multiplies resources and investments of each partner.
Please go to Publications and navigate to Results for final project results.
Images ©Jürgen Biniasch