Piotr Kicki of Poznan University of Technology, Poland

Talk Synopsis

In recent years, we have seen significant progress in the field of robotics, and one of the pillars of this progress is machine learning. Nowadays, we can see that learning aids or even replace robot controllers, motion planners, state estimators, perception pipelines, etc.. However, still, in many cases, we struggle to train effective, safe and generalizable machine learning models, which hampers their applicability in robotics. This seminar aims to address some of these limitations of robot learning by imposing structural priors.

In this presentation, the aforementioned topic will be discussed from two perspectives: (i) learning how to generate motions that satisfy safety and kinodynamic feasibility constraints, and (ii) learning dynamics models for state estimation. The first part will discuss learning how to generate dynamic robot motions that satisfy the constraints and an efficient way to impose boundary conditions on these motions. Finally, the impact of these features on the learning efficiency will be analyzed. The second part will focus on showing how one can impose structure on the learned dynamics model and how it affects its generalization abilities. Moreover, the importance of exploiting the structure of the desired application during the learning phase will be analyzed in the context of learning-based robot state estimation.

These considerations will be supported by experimental analysis performed using challenging and interesting problems, including, autonomous racing with F1/10 car, robotic air hockey and bimanual manipulation.

Biography:

Piotr Kicki is an assistant professor at the Institute of Robotics and Machine Intelligence Poznan University of Technology and research engineer in the Robotics Team at IDEAS NCBR. He received his B.Eng. and M.Sc. degrees in automatic control and robotics from Poznan University of Technology, Poland in 2018 and 2019, respectively. In 2024 he defended his Ph.D. thesis entitled 'Deep reinforcement learning for motion planning in man-made environments'.

Throughout his career, Piotr has contributed to several research projects, including OPUS-LAP INTENTION and H2020 REMODEL. He also completed a doctoral internship with the Jan Peters group at TU Darmstadt. His primary research interests focus on robot motion planning, reinforcement learning, and the application of machine learning in robotics. His work has been published in top-tier venues such as T-RO, RA-L, CoRL, ICRA, and IROS.