Robot Learning
Robot Learning is a wide research field: we employ different forms of Machine Learning in perception (Spatial AI) as well as to learn action: high-level-decisions all the way to fine-grained manipulation. The lab's vision is to employ modern foundation models, large language models, as well as techniques from imitation learning and reinforcement learning to develop truly intelligent mobile robots that scale to diverse and open-ended environments as well as a range of different and complex tasks.
Open-vocabulary scene understanding
In this recent work led by MRL researchers at TUM, we are employing an open-vocabulary and object-centric 3D map representation that directly supports goal-oriented robot exploration of unknown environments. The approach named FindAnything enables 3D segmentation leveraging image foundation models into a coherent 3D representation in real-time – all running on-board a resource-constrained multicopter.
