Perception for Intelligent Mobile Robots
We research mobile robots designed to operate in shared environments for intralogistics applications. To achieve long-term autonomy, we aim to create compact yet comprehensive environment representations with fault awareness. Our approach involves developing map representations that store not only geometric information but also additional details such as semantics and dynamics. We also investigate how radars and other novel sensing techniques can be utilized for localization and mapping.
Mobile robot with radar sensors
Panoptic segmentation and 3D detection
Research Topics
Radar sensor are an emerging sensing modality in mobile robotics. Due to their signal's long wavelength, they are especially robust against visual degradation.
At the chair, we investigate how different types of radar sensors can be used in long-term localization and mapping.
3D perception for autonomous robots operating in outdoor environments is an emerging research field. Common 3D perception tasks include LiDAR Panoptic Segmentation (LPS), which assigns each point in a LiDAR sweep to a semantic category while distinguishing individual instances within relevant categories, and 3D object detection, which localizes objects and estimates their 3D dimensions.
At the Chair, we investigate how multi-sensor fusion can enable a unified 3D perception framework that combines these complementary tasks.
Generative Video Models contain rich priors from internet-scale video data.
We investigate how these models unlock new capabilities in mobile robot navigation.
One of the research directions at the Chair focuses on developing a pipeline to automatically update as-built models of infrastructure worksites. The proposed approach integrates data collected by a sensor box mounted on a heavy-duty machine to generate a precise, reliable 3D semantic map of the environment, which is then used to automatically produce an up-to-date as-built model with minimal human intervention.
Research Projects
Mobile working machines (MWM) are ubiquitous in mining, construction, harbours, forestry, and farming, and without doubt vital for economic and industrial advancement of our society and as such key to sustainability and productivity of our society. Ambitious goals for sustainability, productivity, and safety, coupled with a shortage of human resources, have prompted the European mobile working machines ecosystem to come together in proposing this innovative network of Industrial Doctorates. ENGAGE network will train a new generation of engineers to build next generation mobile working machines for green and autonomous operations. Key European machine manufacturers (Volvo AS, Liebherr, Ponsse, AGCO), and technology suppliers (Bosch Rexroth, Bosch, Novatron, Algoryx, GIM Robotics) have joined forces with 6 academic groups (Tampere University, Karlsruhe Institute of Technology, Umeå University, Aalto University, FZI Forschungszentrum Informatik, Technical University of Munich) highly specialised in AI, robotics, and mechanics from Finland, Germany, Sweden, and Austria to deliver ENGAGE.
More information: dn-engage.eu
AI systems learn increasingly powerful representations from large amounts of data. Yet robots operating in the physical world encounter changing environments, incomplete information, and situations that cannot all be anticipated during training. In our collaboration, we investigate how AI can move beyond reproducing learned patterns towards systems that can interpret new situations, reason about them, and act accordingly.
At the Chair, we are working in the subproject “Pre-training Robot Models on Unstructured Data”
More information: mirmi.tum.de/en/mirmi/mit-tum-collaboration/