3rd Annual Learning for Dynamics & Control Conference

June 7 – 8, 2021 | ETH Zurich, Switzerland Virtual

Over the next decade, the biggest generator of data is expected to be devices that sense and control the physical world.

The explosion of real-time data that is emerging from the physical world requires a rapprochement of areas such as machine learning, control theory, and optimization. While control theory has been firmly rooted in the tradition of model-based design, the availability and scale of data (both temporal and spatial) will require rethinking the foundations of our discipline. From a machine learning perspective, one of the main challenges going forward is to go beyond pattern recognition and address problems in data-driven control and optimization of dynamical processes. Our overall goal is to create a new community of people who think rigorously across the disciplines, ask new questions, and develop the foundations of this new scientific area.

Following the success of the inaugural Learning for Dynamics and Control (L4DC) workshop at MIT in 2019, and L4DC 2020 in (virtual) Berkeley, we are very happy to welcome you to Europe and ETH virtual Zurich.

Due to the ongoing pandemic, we are forced to hold the L4DC 2021 virtually again. More information regarding the virtual format will be sent to registered participants soon. The registration is closed.

For general questions and inquiries please contact the organizers at l4dc@ethz.ch.

The conference proceedings are available as Volume 144 of the Proceedings of Machine Learning Research. They are already on-line here.

Invited Speakers

Conference Program

All times are stated in Pacific Time (PT). On both days the program starts at 8:00 (17:00 CET) and lasts until 13:30 (22:30 CET) on Monday and 12:30 (21:30 CET) on Tuesday. The conference program can also be downloaded here.


Monday, June 7, 2021

YouTube Live-Stream

8:00 – 8:15 Welcome & Introduction by Organizers
8:15 – 8:45Michael Jordan (U.C. Berkeley)
8:45 – 9:30Oral Presentations 1.A: Reinforcement Learning
9:30 – 9:45Coffee Break
9:45 – 10:30Poster Sessions 1.A & 1.B (Not on YouTube)
10:30 – 11:00Sandra Hirche (T.U. Munich)
11:00 – 11:15Coffee Break
11:15 – 12:00Oral Presentations 1.B: Adversarial Learning
12:00 – 12:30Aude Billard (EPF Lausanne)
12:30 – 13:30Brainstorming on Future L4DCs (Not on YouTube)

Tuesday, June 8, 2021

YouTube Live-Stream

8:00 – 8:05 Welcome to the Second Day
8:05 – 8:35Daniel Lee (Cornell University)
8:35 – 9:35Oral Presentations 2.A: Learning Dynamics
9:35 – 9:45Coffee Break
9:45 – 10:30Poster Sessions 2.A & 2.B (Not on YouTube)
10:30 – 11:30Oral Presentations 2.B: Learning Controllers
11:30 – 11:45Coffee Break
11:45 – 12:15Raffaello D’Andrea (ETH Zurich)
12:15 – 12:30Concluding Remarks

Invited Speakers

Monday, June 7, 2021

Michael Jordan (U.C. Berkeley)

Title: On Dynamics-Informed Blending of Machine Learning and Microeconomics

Abstract: Statistical decisions are often given meaning in the context of other decisions, particularly when there are scarce resources to be shared. Managing such sharing is one of the classical goals of microeconomics, and it is given new relevance in the modern setting of large, human-focused datasets, and in data-analytic contexts such as classifiers and recommendation systems. I’ll discuss several recent projects that aim to explore the interface between machine learning and microeconomics, including the study of exploration-exploitation tradeoffs for bandits that compete over a scarce resource, the use of Langevin-based algorithms for Thompson sampling, leader/follower dynamics in strategic classification, and the robust learning of optimal auctions.

Sandra Hirche (T.U. Munich)

Title: “To Sample or not to Sample?” –
Efficient Online Learning in Closed Loop Control Systems

Abstract: Online learning in closed loop control systems is very attractive because it allows the automated identification of highly nonlinear dynamical systems as well as a fast adaptation to dynamically changing environments.

Yet, depending on the application the data collection and the training of models is costly if not even prohibitive for the following reasons: i) The training is computationally expensive and might compromise real-time performance. ii) The generation of training data often requires costly sensor calibration. iii) In human-in-the-loop systems calibration routines are perceived as inconvenient and burdensome. Due to these costs, data should be carefully selected for learning.

In this talk we will demonstrate that the control task in addition to the underlying system dynamics has a strong influence on the required sample complexity. Employing Bayesian principles, we explore methods to quantify epistemic uncertainty with respect to control objectives and how they can be exploited to achieve a high sample efficiency for learning in the closed loop system. Additionally, approaches for efficient non-parametric online learning are investigated to allow the application of the presented methods under real-time constraints.

Aude Billard (EPF Lausanne)

Title: Learning closed-form control laws for robots to react in milliseconds

Abstract: We want autonomous cars, wheelchairs and other mobility devices to transport us autonomously with limited human intervention. We want robots in our homes to cook, clean and entertain us. We want robots on our body to replace a lost limb or to augment our capabilities. All the robots listed above share one common challenge: they must cope with unexpected changes in their environment. For instance, an intelligent wheelchair will have to negotiate its path in a crowd without hitting pedestrians. It will have to do so, while avoiding to brake too abruptly for risk of letting its user fly over. How can the robot recompute a path within milliseconds at time critical situations?

On-line reactivity is not just an issue of having enough CPU on-board of the robot. It requires inherently robust control laws that can re-plan in milliseconds. Most importantly, it needs control laws that are ensured to provide feasible solutions. This talk will show how we can learn a manifold of feasible motions that can be expressed in closed-form, hence ensuring speedy retrieval. We show how traditional regression and classification optimization can be rephrased to provide guarantees on the learned dynamics. The talk will show application of this to learn the nonlinear dynamics of flying objects so as to catch these in flight, and to learn high-dimensional representation of the robot’s joint workspace for whole-body real-time obstacle avoidance. Lastly, a few examples of applications for learning variable force and impedance control for safe human-robot interaction and for safe navigation with a wheelchair travelling in a heavy crowd of pedestrians.

Tuesday, June 8, 2021

Daniel Lee (Cornell University)

Title: Learning for Robot Perception, Planning and Control

Abstract: Conventional computational architectures for robotics segregate representations and processing modules for perception, planning and control. Recent advances in deep learning have shown success in applying end-to-end approaches to robot learning but require large amounts of expensive training data. In this talk, I will contrast these two approaches and present some recent work on statistical bounds in learning-enabled modules and hybrid computational architectures for robot learning.

Raffaello D’Andrea (ETH Zurich)

Title: Optimal Information Gathering for Optimal Decision Making

Abstract: A discussion on the role of flying sensors in the supply chain.

Oral Presentations

Monday, June 7, 2021

Oral Presentations 1.A: Reinforcement Learning (3x 15min)

Chenyu Liu, Yan Zhang, Yi Shen and Michael Zavlanos, Learning without Knowing: Unobserved Context in Continuous Transfer Reinforcement Learning

Rafael Rafailov, Tianhe Yu, Aravind Rajeswaran and Chelsea Finn, Offline Reinforcement Learning from Images with Latent Space Models

Brandon Amos, Samuel Stanton, Denis Yarats and Andrew Gordon Wilson, On the model-based stochastic value gradient for continuous reinforcement learning

Oral Presentations 1.B: Adversarial Learning (3x 15min)

Benoit Landry, Hongkai Dai and Marco Pavone, SEAGuL: Sample Efficient Adversarially Guided Learning of Value Functions

Anshuka Rangi, Mohammad Javad Khojasteh and Massimo Franceschetti, Learning-based attacks in Cyber-Physical Systems: Exploration, Detection, and Cost trade-offs

Udaya Ghai, David Snyder, Anirudha Majumdar and Elad Hazan, Generating Adversarial Disturbances for Controller Verification

Tuesday, June 8, 2021

Oral Presentations 2.A: Learning Dynamics (4x 15min)

Anas Makdesi, Antoine Girard and Laurent Fribourg, Data-Driven Abstraction of Monotone Systems

Guillaume O. Berger, Raphaël M. Jungers and Zheming Wang, Chance-constrained quasi-convex optimization with application to data-driven switched systems control

Sarah Dean and Benjamin Recht, Certainty Equivalent Perception-Based Control

Andrea Sassella, Valentina Breschi and Simone Formentin, Data-driven design of switching reference governors: theory and brake-by-wire application

Oral Presentations 2.B: Learning Controllers (4x 15min)

Yang Zheng, Yujie Tang and Na Li, Analysis of the Optimization Landscape of Linear Quadratic Gaussian (LQG) Control

Fernando Gama and Somayeh Sojoudi, Graph Neural Networks for Distributed Linear-Quadratic Control

Nicholas Boffi, Stephen Tu and Jean-Jacques Slotine, Regret Bounds for Adaptive Nonlinear Control

Anders Rantzer, Minimax Adaptive Control for a Finite Set of Linear Systems

Call for Papers

We invite submissions of short papers addressing topics including:

  • Foundations of learning of dynamics models
  • System identification
  • Optimization for machine learning
  • Data-driven optimization for dynamical systems
  • Distributed learning over distributed systems
  • Reinforcement learning for physical systems
  • Safe reinforcement learning and safe adaptive control
  • Statistical learning for dynamical and control systems
  • Bridging model-based and learning-based dynamical and control systems
  • Physics-constrained learning
  • Physical learning in dynamical and control systems applications in robotics, autonomy, transportation systems, cognitive systems, neuroscience, etc.

While the conference is open to any topic on the interface between machine learning, control, optimization and related areas, its primary goal is to address scientific and application challenges in real-time physical processes modeled by dynamical or control systems.


  • All accepted papers will be presented as posters at this conference. A selected set papers deemed particularly exceptional by the program committee will be presented as oral talks.
  • At least one of each paper’s authors should be present at the virtual conference to present the work. We will inform the authors regarding the logistics of the virtual format in due time.
  • Accepted papers will be published electronically in the Proceedings of Machine Learning Research (PMLR). The authors of accepted papers will have the option of opting-out of the proceedings in favor of a 1-page extended abstract. The full paper reviewed will then be placed on the arXiv repository, but not indexed by the conference.

Dual Submissions

Submissions that are substantially similar to papers that have been previously published, accepted for publication, or submitted in parallel to other peer-reviewed conferences with proceedings or journals may not be submitted to L4DC.

Important Dates

  • Paper submission deadline: November 13, 2020 Extended to November 20, 2020, 5:00 PM EST.
  • Author notification: Early March, 2021.
  • Final Paper Upload Deadline: April 30, 2021. Authors will be directly contacted by e-mail regarding the submission procedure.
  • Final Poster Upload Deadline: May 7, 2021. Extended to May 17, 2021. Authors will be directly contacted by e-mail regarding the submission procedure.
  • Registration: April 26 – June 1, 2021. The registration portal is closed.
  • Conference (virtual): June 7-8, 2021.

Submission instructions

  • Submissions are limited to 10 pages in PMLR format with unlimited allowance for references (Latex Style Sheet). In their submission, authors may point to a tech report if they wish, e.g. on arXiv or on a personal webpage. We updated the template on November 13, 2020 to fix an issue with \includegraphics and on April 13, 2021 to include the proceedings volume number.
  • L4DC reviewing is single blind.
  • We will begin accepting submissions via EasyChair on October 12, 2020.
  • The deadline for submissions is 5:00 PM EST on November 20, 2020.
  • Please contact the program chairs at l4dc@ethz.ch if you have any questions about the policy or technical issues with the submission process.

Confidentiality Policy

We maintain a policy of strict confidentiality for all conference submissions. You must notify your reviewers that they may not disclose or make inappropriate use of the content before publication.


Program Committee

Nikolay Atanasov, University of California San Diego


Peter Bartlett, University of California, Berkeley

Joschka Boedecker, University of Freiburg

Francesco Borrelli, University of California, Berkeley

Valentina Breschi, Politecnico di Milano

Alessandro Chiuso, University of Padova

Samuel Coogan, Georgia Institute of Technology

Stefano Di Cairano, Mitsubishi Electric Research Laboratories

Florian Dorfler, ETH Zurich

Bachir El Khadir, Princeton University

Peyman Esfahani, Delft University of Technology

Mahyar Fazlyab, Johns Hopkins University

Sophie M. Fosson, Politecnico di Torino

Dylan Foster, Cornell University

Simone Garatti, Politecnico di Milano

Konstantinos Gatsis, University of Oxford

Bahman Gharesifard, Queen’s University

Shromona Ghosh, University of California, Berkeley

Sonja Glavaski, PNNL

Shixiang Gu, University of Cambridge

Sofie Haesaert, Eindhoven University of Technology

Hamed Hassani, University of Pennsylvania

Sandra Hirche, Technical University of Munich

Ali Jadbabaie, MIT

Yannis Kevrekidis, Johns Hopkins University

J. Zico Kolter, Carnegie Mellon University

Alec Koppel, U.S. Army Research Laboratory

Danica Kragic, KTH Royal Institute of Technology

Akshay Krishnamurthy, University of Massachusetts, Amherst

Laurent Lessard, Northeastern University

Na Li, Harvard University

John Lygeros, ETH Zurich

Anirudha Majumdar, Princeton University

Kostas Margellos, University of Oxford

Nikolai Matni, University of Pennsylvania

Anastasia Mavrommati, MathWorks

Franziska Meier, Facebook AI Research

Matthias Mueller, Leibniz University Hannover

Gergely Neu, Universitat Pompeu Fabra

Necmiye Ozay, University of Michigan

George Pappas, University of Pennsylvania

Francesca Parise, Cornell University

Pablo Parrilo, MIT

Ioannis Paschalidis, Boston University

Panos Patrinos, Katholieke Universiteit Leuven

Paris Perdikaris, University of Pennsylvania

Maria Prandini, Politecnico di Milano

Victor Preciado, University of Pennsylvania

Maxim Raginsky, University of Illinois at Urbana-Champaign

Anders Rantzer, Lund University

Lillian Ratliff, University of Washington

Benjamin Recht, University of California, Berkeley

Thomas Schön, Uppsala University

Sanjit A. Seshia, University of California, Berkeley

Shahin Shahrampour, Texas A&M University

Milad Siami, Northeastern University

Bartolomeo Stellato, Princeton University

Yuval Tassa, Google/DeepMind

Claire Tomlin, University of California, Berkeley

Sebastian Trimpe, RWTH Aachen University

Kyriakos Vamvoudakis, Georgia Institute of Technology

Yisong Yue, California Institute of Technology

Melanie N. Zeilinger, ETH Zurich