Research

I like working on different kinds of robots, developing new algorithms, testing them on real hardware and to continue advancing our knowledge about motion planning and learning for robotics. Having said that I grow even more excited if those robots have real world applications in helping people, hence I am really keen on working on collaborative and assistive robots. That is why I believe that given my past research experience and future interests I am a good fit for the position as a PhD student with the Robot Learning and Interaction group at the Idiap Research Institute, since with SWITCH and CoLLaboratE they are involved in two projects addressing those kinds of problems.

Trajectory optimization and learning from demonstration have an inherent synergy that can help improve motion planning for robotics. While pursuing a PhD I want to focus my research on discovering the implications of that strong combination of concepts and how it can exploited for different applications. I agree wholeheartedly that robotics could benefit greatly from discovering new techniques for small data problems, since the enhancement of predefined problem solving patterns with skills learned on-the-fly leads to algorithms that can be quickly adapted to changing circumstances.

Embedding task specific constraints that are learned from demonstrations into a trajectory optimization can present a powerful tool, for both approaches bring in their unique strengths. Learning from demonstration can be used to represent high-dimensional knowledge about the task and thus allow for better initial guesses and a smaller search search space for the optimization, thus making it less time consuming and less likely to get stuck in local optima. On the other hand the trajectory optimization can help in generalizing the learned skills.

With PROMPT I already succeeded in proposing the foundation for said combination. PROMPT uses probabilistic movement primitives as the representation for the robot kinematics, which has given me great insight into various learning from demonstration algorithms. However at this point the probabilistic movement primitives are used as the representation of the kinematics model of the robot only, whereas the motion planning problem itself is solved with classical probabilistic methods. This means there is no learning from demonstrations involved yet, thus presenting ample opportunity for extension, most notably the combination with effective motion primitives \cite{loew2020effective}. Which, as I see it, would fit precisely with the research done in the Robotics Learning and Interaction group. \cite{Lembono20RAL} in particular would present a promising collaboration opportunity.

This combination of prewired with learned skills is an extremely intriguing concept for me, since I believe it comes closest to how nature does it. When learning we integrate our experiences from interactions with the world into our innate knowledge and instincts about that world. Therefore in a PhD I want to answer the question of how those instincts can be expressed mathematically and augmented with the experiences a robot makes during its lifetime. In this regard I greatly appreciate the elegance in the approach of combining model predictive control with probabilistic representations of movements in \cite{CalinonLee19}.

Being given the opportunity to help devise the representations that explain our world to robots would be exceptionally exciting for me, since I firmly believe that finding those representations is the key for achieving truly autonomous robots and I love the idea of taking inspiration not only from diverse research fields but also from nature. Consequently I think that I could contribute substantially to the research done by the Robot Learning and Interaction group at the Idiap Research Institute.

I focused on advancing motion planning for robotics during my master studies, which led to the development of a new probabilistic motion planning algorithm called PROMPT \cite{loew2020prompt}. Prior to PROMPT I worked on identifying effective motion primitives from human driving data, which was presented at IROS in \cite{loew2020effective}.

My research was done on various types of robots. The hardware platform that I have been working on most closely is a John Deere Gator that has been automated by CSIRO, the research organization that I was engaged at during my master studies. Not only do I love developing algorithms but also implementing them in code and running them on real robots. I used the Gator to show the implementation of a variety of algorithms, e.g. modified versions of TEB and A*/Hybrid-A* and PROMPT. Apart from the Gator I also implemented and supported the motion planning for the skid-steer platforms that were used by CSIRO in the DARPA challenge.

My work has led to one publication at IROS and one that is under review at RAL. I am currently working on an IJRR paper, under the supervision of Dr. Tirthankar Bandyopadhyay. This will not only implement PROMPT on manipulators but also combine PROMPT with effective motion primitives, leading to a novel framework that combines probabilistic motion planning with learning from demonstrations.

Previous to my engagement at CSIRO I worked at the Robotics Systems Lab at ETH Zurich in the focus project ATHLAS, which aimed to develop a legged landing system for helicopters, which was presented in a paper at IROS \cite{stolz2018adaptive}. My responsibilities were integrating the actuators and writing the control software. We explored two different kinds of actuators, DC motors and fluidic muscles. Latter was the focus of my bachelor's thesis.