Inhalt des Dokuments
I am generally very curious about the thing we call intelligence. What is it and, more importantly, how does it work? In robotics, we want to give answers to these questions through a process of creating and studying robots: machines that exhibit some form of intelligent behaviour. We think that to understand intelligence, it is crucial to study the interaction between robots and our physical world. The facet of intelligence that I find most fascinating is learning because this is what enables robots to adapt to their environment.
Therefore, my research focuses on how robots can learn from experience. Ultimately, they should be able to learn highly complex behaviour from very limited amounts of data, just as humans do. However, I doubt that this is possible tabula rasa (from a blank slate). Instead, I am convinced that robots need to already have prior knowledge about the world incorporated in their learning system.
The questions I am trying to answer in my research are: Which prior knowledge makes efficient robot learning possible? How can we incorporate this knowledge into machine learning? A very promising research direction for this purpose is representation learning: learning to extract the right information from sensory input and representing it in the most useful way. Prior knowledge about the world can be used to learn such representations. These learned representations can dramatically boost the ability of robots to learn new behaviour (see publications below for details).
- 10/2012 - present
Research associate and PhD student at RBO, TU Berlin (Oliver Brock).
Taught courses: Fundamentals of Robotics, Robotics, Advanced Robotics, Robotics Seminar, Robotics Project, Algorithms and Datastructures
- 01/2017 - 04/2017
PhD Intern at DeepMind, London.
- 01/2017 - 04/2017
- 10/2007 - 09/2012
Study of computer science at FU Berlin.
Master of Science (grade: 1.0, major: robotics/AI, minor: psychology).
Thesis: New Approaches to Temporal Abstraction in Hierarchical Reinforcement Learning.
- 01/2012 - 09/2012
Research assistant at MLR, FU Berlin (Marc Toussaint).
Study abroad at UNSW, Sydney.
- 10/2009 - 09/2011
Teaching assistant at FU Berlin.
Taught courses: Functional Programming, Object-Oriented Programming, Computer Science and Society, Software Engineering
Bachelor of Science (grade: 1.4, major: computer science, minor: philosophy).
Thesis: Control of autonomous humanoid soccer robots with XABSL.
- 2008 - 2011
Member of RoboCup team FUmanoids (Raul Rojas).
Awards, Prizes, Scholarships
- 2016 Best Paper Award finalist at IROS
- Rico Jonschkowski, Clemens Eppner, Sebastian Höfer, Roberto Martín-Martín, and Oliver Brock. Probabilistic Multi-Class Segmentation for the Amazon Picking Challenge. IROS, 2016. [VIDEO]
- 2016 Best Systems Paper Award at RSS.
- Clemens Eppner, Sebastian Höfer, Rico Jonschkowski, Roberto Martín-Martín, Arne Sieverling, Vincent Wall, and Oliver Brock. Lessons from the Amazon Picking Challenge: Four Aspects of Building Robotic Systems (PDF, 3,9 MB). Robotics: Science and Systems (RSS), 2016. [INTERVIEW]
- 2015 Winner of the Amazon Picking Challenge at ICRA15 [VIDEO LINKS: APC, OUR RUN].
- 2015 AAAI-15 Robotics Fellowship
- 2011 PROMOS scholarship from FU Berlin
- 2011 4th place RoboCup Worldcup, 2nd place RoboCup German Open,
1st place Technical Challenge @ RoboCup German Open, 1st place RoboCup Iran Open
- 2010 2nd place RoboCup Worldcup, 1st place Technical Challenge @ RoboCup Worldcup, 1st place RoboCup Iran Open
- 2008 2nd place RoboCup German Open
[VIDEO LINK] Lessons from the Amazon Picking Challenge: Four Aspects of Building Robotic Systems. Robotics: Science and Systems, June 2016.
[VIDEO LINK] State Representation Learning in Robotics: Using Prior Knowledge about Physical Interaction. Robotics: Science and Systems, July 2014.
Code for "Learning State Representations with Robotic Priors": github.com/tu-rbo/learning-state-representations-with-robotic-priors
Code for "Probabilistic Multi-Class Segmentation for the Amazon Picking Challenge": gitlab.tubit.tu-berlin.de/rbo-lab/rbo-apc-object-segmentation
Code for "Learning with Side Information": https://github.com/tu-rbo/concarne
Zusatzinformationen / Extras
Tel. +49 30 314-73 110
Fax +49 30 314-21 116
Raum MAR 5.032
Sekretariat MAR 5-1