Bosch Center for Artificial Intelligence

The AutoML Lab led by Prof. Frank Hutter is a joint research lab established by the University of Freiburg and Bosch in order to strengthen joint research and information exchange in fundamentals of Machine Learning. The partnership focuses on the development of Bosch-relevant methods to support the application of AutoML-based solutions in Bosch manufacturing.

• Start: December 2018
• Location: Freiburg, Germany
• Business applications: AutoML
• Research topics: multifidelity optimization, neural architecture search, meta-learning, AutoML for classical ML and AutoML for reinforcement learning

Researchers at TU Wien and BCAI researchers work together on declarative problem solving for combinatorial problems, as well as combinations of symbolic and machine learning-driven AI. The partnership aims at joint research and information exchange as well as the transfer of applications in existing projects at Bosch.

• Start: January 2020
• Location: Vienna, Austria
• Business applications: scheduling and planning in manufacturing
• Research topics: advanced context-based reasoning for knowledge-enhanced combinatorial problems in Bosch manufacturing

Researchers of CMU and Bosch work together in order to conduct cutting-edge research in the field of robust Deep Learning. The goal is to provide better understanding and performance complementing the Bosch research portfolio in robust, safe, and explainable AI.

• Start: April 2018
• Location: Pittsburgh, USA
• Business applications: automated driving, optical inspection, manufacturing
• Research topics: robust deep learning, fundamentals of deep learning, hybrid modelling

The collaboration with the University of Tübingen is established via the Bosch Industry-on-Campus (IoC) Lab, which is located in the campus of the University of Tübingen and is a part of the Cyber Valley ecosystem. The aim is to create new cooperation projects with the research talent of the University of Tübingen and BCAI at the interface between basic academia research and application-oriented research. Currently, the Bosch IoC Lab targets two research fields of Data-Efficient Deep Learning and Safe Deep Learning, with applications to automated driving and driver assistance systems, robotics, and manufacturing.

• Start: November 2018
• Location: Tübingen, Germany
• Business applications: automated driving, driver assistance systems and manufacturing
• Research topics: synthetic data generation and augmentation, few-shot learning, generative modeling, out-of-distribution detection, uncertainty estimation, domain adaptation and generalization

The Tsinghua-Bosch Joint Research Center on Machine Learning is a joint lab established by the University of Tsinghua and Bosch in order to strengthen joint research and information exchange in fundamentals of Machine Learning for manufacturing at Bosch.

• Start: March 2020
• Location: Peking, China
• Business applications: Optimization in manufacturing, highly automated driving
• Research topics: pobabilistic deep learning, explainable machine learning, meta-learning, neuro-symbolic AI, NLP (Natural Language Processing)

SIRIUS is a Norwegian center of research excellence and a hub for academic-industrial collaboration with the focus on AI for industrial digitalization and automation. As a SIRIUS member, BCAI strengthens its research and industrial network and participates in mutually beneficial information exchange.

• Start: January 2020
• Location: Oslo, Norway
• Business applications: factory automation, knowledge analytics
• Research topics: KRR, semantic technologies, digital twins, knowledge graphs, ontologies

Cyber Valley is Europe‘s largest AI consortium including partners from science and industry for research and industrial collaboration in the fundamentals of AI for Automotive, Industry 4.0 and E-commerce. The main goal is to bring together key players from science and industry to advance AI and strengthen the knowledge transfer between collaboration partners.

• Start: January 2018
• Location: Tübingen, Germany
• Business applications: Manufacturing, automated driving, engineering development process
• Research topics: fundamental research covering all aspects of machine learning and computer vision

Since the end of 2016, the University of Amsterdam and the Bosch Center for Artificial Intelligence have been working and researching on deep learning in the jointly established UvA-Bosch Delta Lab. The aim of the research alliance is regular professional exchange and knowledge transfer between science and industry. Beginning 2022, the long-standing cooperation was successfully extended. Delta Lab 2 will focus on research in the area of machine learning for applications in computer vision, generative models and casual learning.

• Start: November 2016, extension in March 2022
• Location: Amsterdam, Netherlands
• Team: ten PhD students, one postdoc, one lab manager
• Research topics: artificial intelligence and machine learning for applications in computer vision, generative models and causal learning
• Business applications: automated driving, computer vision, embedded AI

June 2025

Abstract

With the rapid advancement of artificial intelligence, questions around safety and security are becoming increasingly critical. Various classes of risks emerge — ranging from immediate operational concerns to long-term societal impacts. Understanding and addressing these risks is essential for ensuring responsible development and deployment. Key areas of focus include identifying which risks are most pressing in specific contexts, such as industrial applications, and advancing research that contributes to the creation of robust, secure, and trustworthy AI systems.

June 2025

Abstract

Inertial sensing and sensor fusion are key to navigation across platforms from consumer devices to autonomous systems. This talk presents hybrid neural approaches that enhance nonlinear navigation filtering through deep learning. We introduce a Kalman-informed transformer to adaptively estimate process noise, improving robustness and accuracy in real-world conditions. Applications include visual-inertial odometry and mobile robotics, where a two-stage learning framework uses road curvature to estimate trajectory uncertainty. We also explore neural-assisted measurement updates, integrating learned signals like Doppler velocity and displacement into the filter. Validated across drones, underwater vehicles, and mobile robots, these methods enable more accurate, adaptable, and reliable navigation.

May 2025

Abstract

Machine learning must move beyond human-generated data to address the growing demands of scientific and engineering challenges. Simulations of physical systems — ranging from atomic interactions to fluid dynamics — offer rich, structured data grounded in natural laws. Two key strategies enable this shift: adaptive data generation for neural simulators and physics-informed learning that embeds symmetries and conservation laws. These approaches help bridge the gap between simulation-based science and modern ML, leading to models that not only predict complex phenomena but also enhance our understanding of the physical world.

April 2025

Abstract

The computational cost associated with large models is exceptionally high. Efficient training and inference methods, such as low-precision computation and sparsity, introduce approximations to the original computations, which may result in a loss of accuracy. This talk will present the theory of approximate gradient descent, which offers theoretical guarantees regarding the convergence and convergence rate of efficient approximate training methods. Building on this theory, efficient training and inference algorithms will be introduced, developed through two technical approaches: quantization and sparsity. These include acceleration techniques for feedforward neural network computations, attention computation, activation compression, optimizer compression, and communication compression. Additionally, challenges related to training instability that arise during efficient training will be discussed, along with strategies to address these issues.