Accurate perception, which enables explainable reasoning about possible outcomes of the robot-environment interaction, is essential for many fundamental capabilities, such as localization, mapping, planning or control. Despite several successful robotic solutions, a robust perception architecture that would allow for efficient self-supervised adaptation in an unknown environment remains an open research problem. We claim that the main issues that prevent one from building such architectures are the following: (i) Black-box architectures that neglect the physical embodiment of the robot, such as deep NNs, suffer from poor generalization, weak explainability, catastrophic forgetting and the impossibility of being transferred among different robotics platforms and environments; (ii) Consequent sample-inefficient learning requires a huge number of expensive, domain-specific, human-annotated data that cannot be transferred to other domains; (iii) Resulting complex architectures have a vast number of parameters that have to be adapted/learned jointly to avoid undesirable interference among independently tuned components. We aim to address these issues by designing grey-box, explainable, embodiment-aware, and end-to-end differentiable architectures that enable joint self-supervised adaptation from any successively incoming onboard measurements in a probabilistically consistent way.

Group and supervision: Research will be conducted within the Vision and  Robotics Group (https://cyber.felk.cvut.cz/research/groups-teams/vras/) under the supervision of Assoc. Prof. Karel Zimmermann/Prof. Tomas Svoboda. The group has extensive experience with real robotics, such as successful participation in the DARPA SubT challenge (https://robotics.fel.cvut.cz/cras/darpa-subt/) and several state-of-the-art robotics platforms and sensors (https://robotics.fel.cvut.cz/cras/robots/).