Second Max Planck ETH Workshop on Learning Control
8/9 Febuary 2018, Zurich, Switzerland

After a successful first edition in 2015, we are pleased to announce the second workshop on Learning Control within the Max Planck ETH Center for Learning Systems. The workshop will take place February 8-9 2018 at ETH Zurich. We cordially invite all researchers from ETH Zürich and MPI-IS interested in the area of Learning Control to participate and actively contribute to this workshop.

Aims and Scope

The workshop aims at bringing together researchers from ETH Zürich and MPI-IS working in the area of Learning Control in order to create and grow a community of interest within the Max Planck ETH Center for Learning Systems. The workshop will provide a platform to exchange ideas, present current research, discuss challenges for Learning Control, and initiate future research collaborations within the Center.

The topic and scope of this workshop is Learning Control. Albeit not uniquely defined, we understand Learning Control as the rather broad research area that lies at the intersection of Machine Learning and Automatic Control. This includes, but is not limited to reinforcement learning, machine learning for control, data-driven control, adaptive control, dual control, online learning, active learning for control, model learning, and applications of learning control.

The workshop is open to all researchers from MPI-IS and ETH. Nevertheless, applications of external researchers are possible, though priority will be given to researchers from MPI-IS and ETH (depending on available places). We seek to create an informal atmosphere to foster open discussions and exchange of ideas.


All participants are asked to submit a maximum one page abstract detailing their research interests. In addition to presenting research results in the area of Learning Control, we encourage participants to also include open research questions, early ideas, or any topic that can lead to interesting and fruitful discussions in this exciting area.

Submissions will be briefly reviewed to ensure adequacy with the scope of the workshop. A few submissions will be selected for short plenary presentation. All other accepted submissions will be presented in interactive poster format.


In addition to interactive presentations and short talks, there will be invited keynote talks, a panel discussion, as well as social events (dinner on Thursday evening) with ample room for discussions and informal interactions.

We are excited to have the following invited speakers:

  • Thomas Schön, Professor at Department of Information Technology, Uppsala University, Sweden
  • John Lygeros, Professor at Automatic Control Laboratory, ETH Zurich, Switzerland
  • Nicolas Heess , Deepmind

Costs & logistics

Accommodation and travel costs for participants from MPI-IS (CLS members, associated members, students / co-workers of their groups) are covered by the Max Planck ETH Center for Learning Systems.

The following hotels have been booked.

Meals (lunch, dinner, coffee breaks) are offered for all participants.

Workshop location

All workshop activities, including registration on the first day, take place at ETH Zurich, CAB building, lecture room CAB G 51.
ETH Zurich
Department of Computer Science
CAB Building
Universitätsstrasse 6
8092 Zurich

Directions from Zurich main station to CAB building:

Tram no. 6 (direction Zoo) or Tram no. 10 (direction Airport) to stop „ETH/UniversitätsSpital“.

The dinner on Thursday evening Feb 8, 2018 will be at Hotel Uto Kulm on top of Uetliberg (detailed information will be provided during the workshop).

Directions to Uetliberg (dinner):

From ETH/Universitätspital to Zurich Main station take Tram 6 Direction Enge or Tram 10 Direction Bahnhofplatz. From Zurich main station, take the Sihltal Zurich-Uetliberg Bahn SZU (S10) from the underground station on track 21/22. The SZU runs weekdays every half-hour and takes 20 min to get to Uetliberg station. From there it is a 7 minute walk to the hotel.

Trams 6/10 from ETH/Universitätsspital need 6-7 minutes to Zurich main station and run every 5-10 minutes. The train schedule from Zurich main station is as follows:

  • 18.05-18.27
  • 18.35-18.57



  • November 2017 - Registration/abstract submission opens (registration through website below)
  • 15th December 2017 - Registration and abstract submission deadline
  • 23rd December 2017 - Notification of acceptance
  • 8 February 2018 (around noon) - Workshop starts
  • 9 February 2018 (early evening) - Workshop ends


Schedule (talks) & poster sessions

Schedule (overview)

February 8th (Thursday)

Description From To
train from TUe to ZH HB: arrive 10:25/11:25/12:25
Registration 11:00 13:00
Lunch (Standing lunch in the Foyer CAB G 10.005) 12:00 13:00
Opening 13:00 13:15
Invited talk: Thomas Schön: System identification meets the Gaussian process (50 min + Q/A)
Show details
13:15 14:15

Abstract: The Gaussian process (GP) has over the past 5-7 years become a standard tool within system identification. One important starting point was the successful use of the GP in estimating the impulse response of a linear system. Since then the GP has also been used to construct non-parametric nonlinear state-space models and for on-line nonlinear black-box modelling, etc. In this talk I will focus on one of our recent developments where we show how the GP can be used to solve stochastic optimization problems. These problems have been studied for a long time and due to the success of deep learning these problems are now more relevant than ever. Our main motivation in this talk comes from another problem though, namely the problem of nonlinear system identification. Here, the very nature of the problem is such that we can only access the cost function and its derivative via noisy observations, since there are no closed-form expressions available. Via sequential Mote Carlo methods (aka particle filters) we can obtain unbiased estimates of the likelihood function. We start from the fact that many of the existing quasi-Newton algorithms can be formulated as learning algorithms, capable of learning local models of the cost functions. Inspired by this we can start assembling stochastic Newton-type algorithms, applicable in situations where we only have access to noisy observations of the cost function and its derivatives. This is where our interest lies. We will show how we can make use of the GP model to learn the Hessian allowing for efficient solution of these stochastic optimisation problems. I will also mention some ongoing work where we scale this to much higher dimensions (in terms of the size of the dataset and the number of unknown parameters). If there is time towards the end I will also show a few additional GP constructions that we have been developing recently.


Poster session / coffee - See poster sessions below 14:15 16:00
Participant talks: 16:00 17:00
Matteo Turchetta
Goal Oriented Safe Exploration in Discrete Markov Decision Processes
Co-Authors: Felix Berkenkamp, Andreas Krause
16:00 16:15
Andreas Doerr
Robust Learning of dynamics models for model-based policy search
Co-Authors: Christian Daniel, Duy Nguyen-Tuong, Alonso Marco, Stefan Schaal, Marc Toussaint, Sebastian Trimpe
16:20 16:35
Ruben Grandia
Model learning for legged robots
Co-Authors: Marco Hutter
16:40 16:55
Travel Uetliberg (Meeting point: CAB Building Registration area) 17:30 18:30
Dinner (Hotel Utokulm, Uetliberg) 19:00 23:00

February 9th (Friday)

Description From To
Invited talk: Nicolas Heess: Deep reinforcement learning for control -- algorithms and architectures
Show details
09:00 10:00


Enabling an embodied agent to act autonomously in the physical world is one of the long-standing challenges of artificial intelligence. Besides its practical implications it is also an interesting testbed for generallearning algorithms.

Deep learning based approaches have recently achieved impressive results in domains such as Atari and Go, learning sophisticated behaviors from scratch. They hold equal promise for learning solutions for difficult motor control tasks. Such tasks can, however, pose additional challenges, especially when working with real robotics hardware. In my talk I will describe some of these challenges, and I will present an overview of deep reinforcement learning algorithms and architectures for learning control, discuss some of their strengths and weaknesses, and show some applications both in simulation and in the real world.

Poster session / coffee - See poster sessions below 10:00 11:50
Participant talks: 11:50 12:30
Arash Mehrjou
Controlled Generative Adversarial Networks
Co-Authors: Bernhard Schölkopf, Saeed Saremi
11:50 12:05
Sebastian Blaes
Using Embodied Exploration for Reinforcement Learning
Co-Authors: Jia-Jie Zhu, Georg Martius
12:10 12:25
Lunch (Standing lunch in the Foyer CAB G 10.005) 12:30 14:00
Invited talk: John Lygeros:A statistical learning perspective on scenario optimisation (50 min + Q/A)
Show details
14:00 15:00

Abstract: Though seemingly very different, randomised optimisation and statistical learning theory in fact enjoy deep connections. One of these relies on the notion of compression. This is the observation that under some conditions, when using a large data sample to make decisions (learn a "concept" in machine learning, or minimise a cost function subject to constraints in optimisation), the final decision we come to will generally only depend on a small sub-sample of fixed cardinality. The effect of the remaining samples is only to provide confidence; the more samples we have seen the more confident we are that the decision we made is good. The talk will introduce the compression notion using learning of concepts in statistical learning theory. We will then explore its implications for decision making under uncertainty based on scenario optimisation.


Participant talks: 15:00 16:00
Daniel Kappler
Increasing Sample-Efficiency via Online Meta-Learning
Co-Authors: Stefan Schaal and Franziska Meier
15:00 15:15
Kim Peter Wabersich
Model predictive safety certificates from data for learning-based control
Co-Authors: Melanie N. Zeilinger
15:20 15:35
Carmelo Sferrazza
Iterative learning for the generation and tracking of trajectories using parametrized model predictive control
Co-Authors: Michael Muehlebach, Raffaello D’Andrea
15:40 15:55
Closing 16:00 16:15

Poster Sessions

Febuary 8th (Thursday)

Author Title Co-authors
Carl Jidling Linearly constrained Gaussian processes Niklas Wahlström, Adrian Wills, Thomas B. Schön Abstract
Sebastian Curi Ph. D. Research goals Robust reinforcement learning Abstract
Dominik Baumann Learning to Save Communication Friedrich Solowjow, Sebastian Trimpe Abstract
Aravind Elanjimattathil Vijayan Approaches to solving simultaneous control and locomotion problem for quadrupeds with manipulators Abstract
Markus Giftthaler The ‘Control Toolbox’ - An Open-Source C++ Library for Robot Modelling, Control, Estimation and Learning Michael Neunert, Jonas Buchli Abstract
Ashish Cherukuri Data-driven distributed optimization for multiagent systems Abstract
Oleksandr Zlatov Implementation and investigation of some state-of-the-art deep reinforcement learning approaches Sebastian Trimpe Abstract
David Hoeller Reinforcement Learning in Simulation (doing PhD with Marco Hutter, remark by MS) Abstract
Samuel Bustamante Continuous control of a robotic arm with a non-invasive brain-machine interface Moritz Grosse-Wentrup Abstract
Jan Carius Nonlinear Optimal Control for Switched Systems Farbod Farshidian Abstract
Alexander von Rohr Learning Control for Adaptive Locomotion of Soft Microrobots Stefano Palagi, Sebastian Trimpe Abstract
Mohammad Khosravi Learning the Energy Consumption in Buildings: A Joint Classification and Nonlinear Regression Algorithm Annika Eichler, Roy Smith Abstract
Stefan Stevsic Learning Robot Control from End-user Specifications Manuel Kaufmann Abstract
Jia-Jie Zhu Guiding Meta Reinforcement Learning using Optimal Control Georg Martius, MPI-IS Abstract
Julian Zilly A spectral learning theory for robotics Andrea Censi, Jacopo Tani, Emilio Frazzoli Abstract
Lukas Hewing Cautious Nonlinear Model Predictive Control with Gaussian Process Dynamics Melanie N. Zeilinger Abstract
Angeliki Kamoutsi Data-driven approximate dynamic programming: A linear programming approach Tobias Sutter, Angeliki Kamoutsi, Peyman Mohajerin Esfahani, and John Lygeros Abstract
Matthias Hofer Learning Control for Soft Robotic Manipulator Abstract
Elena Arcari Uncertainty Learning and Control of a Robotic Arm Elena Arcari, Andrea Carron, Lukas Hewing, Markus Giftthaler, Melanie N. Zeilinger Abstract
Luca Perrozzi Deep learning for jet reconstruction in the CMS experiment Elena Arcari, Andrea Carron, Lukas Hewing, Markus Giftthaler, Melanie N. Zeilinger Abstract

Febuary 9th (Friday)

Author Title Co-authors
Steve Heim Designing Natural Dynamics that are Easy to Exploit Abstract
Danny Driess Learning to Control Redundant Musculoskeletal Systems Daniel Hennes, Marc Toussaint, Syn Schmitt Abstract
Konstantinos Kokkalis Locally Weighted Learning Control with Stability Guarantees Konstantinos Kokkalis, Sebastian Trimpe Abstract
Okan Koc Optimizing Robot Striking Movements with Iterative Learning Control Guilherme Maeda, Jan Peters Abstract
Vinay Jayaram Reinforcement learning for prosthetic control Abstract
Vassilios Tsounis Hierarchical Reinforcement Learning for Hybrid Control in Legged Locomotion Abstract
Dengxin Dai PathTrack: Fast Trajectory Annotation with Path Supervision Santiago Manen, Michael Gygli, Dengxin Dai, Luc Van Gool Abstract
Xiaoguang Dong Planning Spin-Walking Locomotion for Automatic Grasping of Microobjects by An Untethered Magnetic Microgripper Metin Sitti Abstract
Zahra Grosser Vibration Detection and Suppression for Distributed Mechanical Systems Melanie N Zeilinger Abstract
Andrea Carron Safe Learning for Distributed Systems Abstract
Marcel Menner Personalizing Human-in-the-Loop Control Systems Melanie Zeilinger Abstract
Michal Rolínek Extrapolation via Learning Physics G. Martius, F. Solowjow, S. Trimpe Abstract
Johannes Kirschner Stochastic Bandits with Heteroscedastic Noise Andreas Krause Abstract
Cristina Pinneri Self-organized discovery of goal-free behaviors: a new way to interactive robotics Georg Martius Abstract
Edgar Klenske The Spectrum of Dual Control Abstract
Angeliki Kamoutsi A scenario based approach to data-driven inverse stochastic optimal control Angeliki kamoutsi, Tobias Sutter, John Lygeros Abstract
Manuel Wuthrich Approximation Maximization - a Novel Policy Optimization Algorithm Alexander Herzog, Stefan Schaal Abstract
Alonso Marco On the Design of LQR Kernels for Efficient Controller Learning Philipp Hennig, Stefan Schaal, and Sebastian Trimpe Abstract
Lukas Fröhlich Active Model Learning for Feedback Controlled Dynamic Systems Edgar Klenske, Melanie N. Zeilinger Abstract

Information for presenters

Poster presentations:

Please bring an A0 poster (portrait).

Participants talks:

We have scheduled 20 min for each talk. Please prepare a 15 min presentation and allow 5 min for questions and discussions afterward.


Event registration is now closed



Sebastian Trimpe

MPI-IS, Autonomous Motion Department

Melanie Zeilinger

ETH Zürich, Institute for Dynamic Systems and Control

Georg Martius

MPI-IS, Autonomous Learning Group