Autonomy@Scale

Highly automated vehicles are an integral part of everyday traffic and will be used even more in the future with further autonomy. The world is dynamically evolving with an "open-world" nature, creating new domains for highly/fully automated vehicles to learn. The environmental changes and developments are not only important for in time and location aspects, but also the constant developments in the field of hardware, such as sensors, software, and development states, pose great challenges. Up until now, artificial intelligence (AI) modules in autonomous driving applications are only scalable to a limited extent. The reason lies in the training strategies applied, in which algorithms have to be trained for each new domain. This results in enormous development costs.

This workshop addresses the need for new training methods for AI systems and researches. Disruptive methods of "effective" machine learning can enable a more efficient and unrestricted use of AI and thus “scaling of autonomy” in new domains. With this approach, automated vehicles can access new markets faster and respond agile to new demands.

The scaling of autonomy can be achieved by using the gained knowledge from known domains and focusing on learning gaps to the target domain. This is also called Delta Learning, which can be achieved by applying methods from transfer learning and incremental learning.

The focus of this workshop is on methods and approaches that contribute to the scaling of autonomy solutions (Autonomy@Scale). The goal of these methods is to reduce the dependency on high data volumes, to use different data sources and also to be able to use knowledge that has already been gained from other domains:

• Semi-Supervised, Weakly-Supervised Learning and Unsupervised Learning
• Few/Single Shot Learning
• Multi-Task Learning
• Active Learning
• Continuous Learning and Incremental Learning
• Knowledge Transfer and Knowledge Distillation
• Transfer Learning
• Domain and Task Adaptation
• Uncertainty Estimation and Out of Distribution Detection
• Generative Modelling (e.g. GANs, VAEs)
• Relational Learning (e.g. Graph Networks, Self-Attention)

The application areas for the above methods are different sources of domain shifts and changes. These changes can be caused by:

• Changes in task or label set (e.g. new classes or class imbalance)
• Changes from simulation to real environment and vice versa (sim2real)
• Cross sensor adaptation and using the knowledge from other sensors (with different technology, type, or position and orientation)
• Place and locations such as changes in road domain (e.g. highway to urban) or country (e.g. USA to china) or traffic situations (e.g roundabout to intersection)
• Changes caused by time such as seeing new classes (new mobility concepts such as escooters) or changes in weather or lighting conditions.

The application areas may be applied to different sensor technologies (e.g. camera, LiDAR, and RADAR) and also used in different functional modules such as perception, fusion, behavior and trajectory planning.

Coming soon...

The workshop is organized by members of the German funded project of Delta Learning (www.ki-deltalearning.de/en) together with great industry partners such as Mercedes-Benz, BMW, VW, Bosch and ZF.

Organizer:

• Larissa Triess from Mercedes-Benz AG
• Prof. Cristobal Curio from University of Reutlingen
• Dr.-Ing. Saqib Bukhari from ZF