Postdoc Position - End-to-end learning with explainability. Self-supervised learning (01-Postdoc-Svoboda)

Research group Vision for Robotics and Autonomous Systems (VRAS) at the Department of Cybernetics, Faculty of Electrical Engineering, Czech Technical University in Prague, is looking for outstanding postdoctoral researchers. 

In the context of rescue robotics scenarios - mobile robots entering unstructured, possibly hostile environments - the successful candidate will research end-to-end fully differentiable pipelines consisting of not only neural network modules but also other blocks containing implicit optimisation such as factor graph-based SLAM or optimal control. Once constructed, such an architecture can be trained from the sensory data on the input (e.g., cameras, LIDAR, IMU, encoders) over localisation, terrain traversability estimation, to motor commands (actions), in an end-to-end manner, using the error backpropagation algorithm. 

While collecting vast amounts of data from various onboard sensors on contemporary vehicles is easy, providing manual annotations for supervised learning is extremely challenging and consequently expensive. It is especially appealing in safety-critical applications of autonomous driving, where performance requirements are extreme and datasets are large. We leverage the strong prior knowledge about various semantic classes that appear in traffic scenarios in order to construct novel self-supervised losses that allow learning from these unlabeled data. These losses typically encode first-principle models (learning the 3D flow and terrain models from sequential data or learning the 3D shapes of objects) or behavioral patterns induced by traffic rules (learning to predict future motion).  

Group and supervision: Research will be conducted with the Vision for Robotics and Autonomous Systems group (https://cyber.felk.cvut.cz/research/groups-teams/vras/), supervised by Profs. Tomas Svoboda and Karel Zimmermann. The group was leading the successful participation of the CTU team in the DARPA Subterranean Challenge (3rd overall and 1st among self-funded teams in Tunnel Circuit (2019) and Urban Circuit (2020); 2nd overall in the Virtual challenge of the Final Event (2021).

Hardware infrastructure:  wheeled, tracked, and legged robots (including 2 Spots and 2 Spots with arm) for search and rescue operations.