Student Projects

In some applications, a controller has to be designed online for a system that is a priori unknown. A medical ventilator should for instance work directly on any patient without prior tuning. It should safely adapt to the patient’s morphology in real time while pumping air to/from the lungs. Problems of this kind fall under the umbrella of Adaptive Control.

The goal of this project is to design an adaptive controller for a relevant application of your choice. Possibilities include:

– Medical ventilator, in collaboration with Hamilton Medical,

– Autonomous vehicle, as part of the challenge proposed by Comma.AI.

Expected deliverables:

– Design of an adaptive controller,

– (If relevant/you’re interested) Provide theoretical guarantees about your control scheme, e.g. stability, constraint satisfaction, regret,

– Numerical experiments.

Required knowledge and tools:

– Control theory

– (ideally) System identification

– (ideally) Convex optimization

– Python

Comment

If interested, please send your transcript to Sabri El Amrani ([email protected]).

Professor(s)

Giancarlo Ferrari Trecate

Administration

Barbara Marie-Louise Frédérique Schenkel

The EPFL babyfoot is under continuous improvement.

While the babyfoot can easily intercept the ball and kick toward the opponent goal with success, it can also juggle and shoot if the ball is still.

The difficulty is to capture the ball, make a pass and chain precise actions. This project aims at improving the ball handling while it is moving to permit capture, pass and juggling, and then shoot toward opponents goal.

Students suggested improvements are also welcome.

The babyfoot is programmed using LabVIEW, knowledge of this language is prerequisite for strategy related projects.

Professor(s)

Christophe Salzmann (Laboratoire d’automatique 3)

Site

https://www.epfl.ch/labs/la/pi/babyfoot/

ROJECT SUMMARY

Adaptive optics (AO) systems are critical in overcoming atmospheric distortions in ground-based telescopes, enabling sharper and more detailed astronomical observations. This project focuses on designing and synthesising a robust, real-time controller for an AO system. By leveraging advanced control theory, the project aims to significantly enhance the performance of AO systems, ensuring reliable correction of wavefront distortions and improving the quality of astronomical images.

BACKGROUND AND MOTIVATION

Telescopes observing through Earth’s atmosphere suffer from distortions caused by turbulent air layers, which degrade image resolution. Adaptive optics counteract this by using deformable mirrors and wavefront sensors to correct for distortions in real-time.

Current AO systems, while effective, face limitations in:

* Handling dynamic, unknown atmospheric conditions.

* Scalability for large telescopes and next-generation systems.

By synthesising a controller tailored to these challenges, this project seeks to push the boundaries of AO performance, paving the way for discoveries in astronomy and astrophysics.

OBJECTIVES

1. Design and develop a robust controller that addresses dynamic atmospheric variations.

2. Validate performance through simulation (and hardware) implementation using an AO testbed.

3. Ensure scalability and adaptability of the controller for different telescopes and operational conditions.

REQUIREMENTS

The project demands a solid academic foundation in control courses, particularly in ‘Advanced Control Systems.’ Proficiency in the frequency domain approach and robust control techniques is especially critical.

Professor(s)

Alireza Karimi, Vaibhav Gupta

Administration

Barbara Marie-Louise Frédérique Schenkel

External

Department of Astronomy, UNIGE, Isaac Dinis, [email protected]

Site

ddmac.epfl.ch, la.epfl.ch

Professor(s)

Maryam Kamgarpour (Recherche en systèmes)

Administration

Barbara Marie-Louise Frédérique Schenkel

Site

https://www.epfl.ch/labs/sycamore/imitation-learning-inverse-reinforcement-learning-for-dexterous-robotic-manipulation-master-thesis/

The hybrid microvibration damping platform (MIVIDA) developed at CSEM was designed to demonstrate the capabilities of data-driven

control design methods for adaptive disturbance rejection. The modular platform comprises an adjustable number of passive dampers, a

set of proof mass actuators (PMAs) creating a 6 DoF force tensor, and a payload interface capable of accommodating various types of

sensitive instruments. Using the set of actuators and based on the accelerometer measurements close to the payload, the platform

actively rejects disturbances induced to the suspended base plate using an external inertial shaker. At the current stage, the

implemented data-driven control laws need to be recomputed every time a new payload is mounted. The objective of the master project

is to develop a robust performance controller delivering stability and performance for a variety of payloads without requiring retuning of

the system.

Different control design methods shall be studied in the scope of the project. A data-driven uncertainty set can be acquired by

performing system identification experiments for a number of different payloads. A robust controller can be computed based on the

acquired input/output data using convex optimisation. Further possible methods involve machine learning using the input/output time

series as training dataset. As an example, offline reinforcement learning allows to design recurrent neural networks (RNN) or

physics-informed neural networks (PiNN) enabling robustness of the controller for varying system properties which cannot be modelled.

The exact nature and amount of design methods to be studied will be defined during the project.

The student will be working in the Sensing & Control laboratory at CSEM Neuchâtel in the scope of a PdM in industry.

Professor(s)

Alireza Karimi

External

CSEM Neuchâtel, Elias Klauser, [email protected]

Site

https://www.csem.ch/en/tailored-services/microvibration-testing/

The goal of this project is to design a motion controller to drive a bumper car into a desired position and orientation within a constrained environment. Controlling such a vehicle is particularly challenging due to its non-holonomic nature, which limits its motion capabilities, and the limited maneuvering space available in our laboratory setup.

The student will design a suitable control strategy and test its validity in simulation. If time allows, the controller will also be implemented and tested on the physical prototype in the lab to test performance in real-world scenarios.

The resulting controller will serve as a key component for future automated crash tests and other applications requiring precise control of the bumper car in confined spaces.

Objectives:

[OB] Design a motion controller for the bumper car and test its validity in simulation;

[SO1] Implement the control strategy on the bumper car prototype.

Required knowledge and tools:

– Control theory;

– Experience with simulation environments (Simulink/Python);

– (Optional) Experience with ROS2.

Comment

Students interested in the project should contact Simone Baratto ([email protected]).

Professor(s)

Giancarlo Ferrari Trecate

Administration

Barbara Marie-Louise Frédérique Schenkel

Site

https://www.epfl.ch/labs/la/pi/

Professor(s)

Maryam Kamgarpour (Recherche en systèmes)

Administration

Barbara Marie-Louise Frédérique Schenkel

Site

https://www.epfl.ch/labs/sycamore/imitation-learning-inverse-reinforcement-learning-for-dexterous-robotic-manipulation-master-thesis/

Heterogeneous Monitoring of of Industrial Processes with Graph Neural Networks

The Industrial Internet of Things (IIoT) generates vast quantities of data from interconnected sensors and actuators operating within complex industrial systems. While conventional data-driven methods analyze these signals with temporal context, they often fail to exploit the relational structure of industrial processes, such as physical connections, control interaction and functional dependencies between process variables and equipment.

In industrial plants, sensor data can be broadly categorized into process state variables (e.g., temperature, pressure, flow), actuation or load variables (e.g., motor current, valve position), and equipment health indicators (e.g., vibration, acoustic emission). These variables are tightly coupled: process conditions influence equipment load, which in turn affects mechanical degradation. However, in current industrial practice, process monitoring and equipment condition monitoring are typically treated as separate problems, obscuring these interactions and delaying fault detection.

As a result, many equipment faults are only detected once mechanical damage has already progressed. For example, pumps, among the most failure-prone components in process manufacturing, often experience clogging due to suboptimal process operation, such as inappropriate temperature changes that induce crystallization or solidification. While vibration-based monitoring may detect the issue at a late stage, it often cannot distinguish whether the root cause lies in process conditions or intrinsic equipment failure. Capturing the relationships between process variables, load indicators, and equipment health is therefore essential for early fault detection and meaningful root cause analysis.

Graph Neural Networks (GNNs) provide a powerful framework to model such systems by explicitly representing sensors and actuators as nodes in a graph, with edges encoding their interactions. By leveraging sensor relationship graphs inferred from IIoT data and guided by physical priors and known process dependencies, GNNs can capture structured, directional dependencies that are typically lost in traditional machine learning approaches.

Thesis Goal

This thesis aims to develop a framework for heterogeneous monitoring that jointly models process behavior and equipment health. By combining data-driven graph inference with physics- and process-informed constraints, the proposed approach seeks to enable early fault detection, unsupervised fault diagnosis, and interpretable insights into how process operation affects equipment degradation.

You will work with real industrial IIoT data and addresses a practical and pressing challenge in process manufacturing, conducted as part of an early stage startup project. Ultimately, the goal is to support operators with actionable insights that enables timely and informed interventions to ensure process reliability.

Requirements

– Good knowledge of Python programming.

– Good understanding of fundamental machine learning concepts.

– Experience with GNNs is advantageous.

Professor(s)

Mengjie Zhao (Laboratoire d’automatique 3), Colin Neil Jones

Administration

Barbara Marie-Louise Frédérique Schenkel

External

KIIO

Site

https://sirop.org/app/6bd357eb-d958-4b1c-b40d-ef3ee4d31176?_k=VXrSaqjZ7UDEUApl

Outline

This project aims to apply the theory of signalling in games to local energy markets, with simulations on real-world distribution grids.

Motivation

Local Energy Markets are a framework to integrate small-scale energy resources at the household level towards the wider grid. Local Energy Markets typically consist of a market operator, who is responsible for pricing and information communication, and prosumers (producers+consumers) at the residential scale. The objective of the market operator is to coordinate prosumers for fair and efficient market outcomes. On the other hand, the prosumers may have selfish objectives (profit maximization).

Prosumers typically have stochastic production sources (e.g. solar panels), whose realization depends on the weather and is unknown to the prosumers; and flexible demand (e.g. Electric Vehicle charging), which can be controllable. The profits for prosumers may depend on the combination of their stochastic production and controllable consumption. Prosumers may choose to deploy their demand side flexibility in order to achieve selfish objectives, like profit maximization, which may come at the expense of fairness and efficiency for the larger market.

The goal of the project is to investigate whether the market operator can strategically communicate forecasts of weather and/or renewable production [1] to prosumers in order to achieve more efficient and fair market outcomes. The project will involve applying existing techniques from information design in games [2] and conducting simulations on real-world grid and consumption data.

References

[1] Rene Aid, Anupama Kowli, and Ankur A. Kulkarni. Signalling for Electricity Demand Response: When is Truth Telling Optimal? arXiv:2302.12770. July 2023.

[2] Emir Kamenica and Matthew Gentzkow. “Bayesian Persuasion”. en. In: American Economic Review 101.6 (Oct. 2011), pp. 2590-2615. issn: 0002-8282.

Comment

The project will involve a mix of theory and simulations on real world electricity grids. The candidate should have strong background in coding (cvxpy in python) and familiarity with optimization, game theory and probability

Professor(s)

Maryam Kamgarpour (Recherche en systèmes), Saurabh Dilip Vaishampayan

Administration

Barbara Marie-Louise Frédérique Schenkel

Site

https://www.epfl.ch/labs/sycamore/information-design-in-local-energy-markets/

**Project Overview**

The project explores meta‑learning methods to develop neural‑network controllers that adapt rapidly to changing disturbance profiles, such as varying wind conditions. By leveraging prior experience across tasks, the controller should generalize efficiently to new scenarios with minimal additional data.

**Objectives**

– Review state-of-the-art meta‑learning algorithms (e.g., MAML) and their suitability for control.

– Analyze neural‑network control architectures and identify key design choices for disturbance rejection.

– Implement selected meta‑learning algorithms in Python (using PyTorch) within a simulated quadcopter environment.

– Evaluate adaptation speed and robustness under diverse wind profiles and quantify improvements over baseline controllers.

**Methodology**

– Conduct a concise literature survey on meta‑learning and neural control.

– Select and configure a quadcopter simulator for disturbance testing.

– Train meta‑learned controllers and compare against standard training (from scratch).

**Required Skills & Tools**

– Control Theory

– Machine Learning: Meta‑learning concepts, PyTorch.

– Software: Python, chosen quadcopter simulation framework (e.g., PyBullet).

——————————————————————————————————–

**Key Reference**

– Finn, C., Abbeel, P., & Levine, S. (2017). Model‑Agnostic Meta‑Learning for Fast Adaptation of Deep Networks. In Proceedings of ICML (pp. 1126-1135).

Comment

If you are interested, please send your transcript (BSc and MSc) to the following email:

[email protected]

[email protected]

Professor(s)

Giancarlo Ferrari Trecate

Administration

Barbara Marie-Louise Frédérique Schenkel

The proposed project aims to explore and evaluate the use of meta-learning techniques in control systems.

Meta-learning focuses on enabling models to adapt quickly to new tasks with limited data by leveraging prior experience across related tasks. In this context, the project will focus on implementing algorithms that allow Neural Network-based controllers to efficiently generalize across a variety of control scenarios, adapting to new environments with minimal additional training.

The student will investigate the performance of established meta-learning methods, such as Model-Agnostic Meta-Learning (MAML) [1], in control tasks. By integrating these techniques with control-specific neural architectures [2] and evaluating them in simulation, the project will explore the potential of meta-learning for enhancing adaptability, robustness, and sample efficiency in dynamical systems. Simulations will be conducted on the Franka Robot, linking theoretical methods with practical, real-world application.

Main tasks:

-Review and understand relevant meta-learning algorithms

-Study Neural Network-based approaches to control

-Implement and evaluate selected approaches for control

-Perform simulations with the Franka Robot to assess real-world applicability

Required knowledge and tools:

-Control theory

-Machine learning

-Python and relevant libraries (PyTorch) for neural network implementation.

Reference papers:

– [1] C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” in Proc. 34th Int. Conf. Machine Learning, Sydney, Australia, 2017, pp. 1126-1135.

– [2] L. Furieri, C. L. Galimberti, and G. Ferrari-Trecate, “Learning to boost the performance of stable nonlinear systems,” IEEE Open Journal of Control Systems, p. 342-357, 2024.

Comment

Contact:

[email protected]

[email protected]

Professor(s)

Giancarlo Ferrari Trecate, Christophe Salzmann

Administration

Barbara Marie-Louise Frédérique Schenkel

Previous work in the DECODE Lab (Furieri, 2024) proposed a method for learning stability-preserving nonlinear optimal controllers with neural networks. This framework, referred to as performance boosting (PB), has proven to be powerful for nonlinear and distributed control.

Recently, PB has been extended to switching controllers (Saccani, 2025) to maintain good performance in time-varying environments. While sufficient conditions for stability preservation during switches have been identified, the design of adaptive controllers for good practical performance remains an open challenge.

The goal of this project is to draw inspiration from online convex optimization (Hazan, 2016) to design new, theoretically sound, switching schemes for online adaptation. These schemes will have to be implemented and rigorously tested in a simulated setting for the control of a wind turbine.

Objectives:

[OB] Design new online switching schemes for performance-boosting controllers;

[SO1] Prove regret guarantees for the proposed switching schemes.

[SO2] Implement the proposed solutions in a simulated setting.

Required knowledge and tools:

– Control theory;

– Convex optimization;

– (Strong) experience with both Python and PyTorch.

Comment

Interested students should contact Sabri El Amrani ([email protected]) and Julien Pallage ([email protected]).

Professor(s)

Giancarlo Ferrari Trecate, Sabri El Amrani

Administration

Barbara Marie-Louise Frédérique Schenkel

External

DECODE

The objective of this project is to design a controller for a nonlinear system, a 2DoF Gyoscope. The nonlinearities in the dynamics will be handled using appropriate methods like feedback linearisation. Uncertainty quantification techniques using data will then be employed to design controllers that robustly stabilise the system with respect to parameter uncertainties and unmodelled nonlinearities. The designed controller should then be validated on the system. Controller implementation and real-time interface would be done through LabVIEW.

Comment

Knowledge of courses System Identification and Advanced Control Systems are ideal. Contact Vishnu Varadan ([email protected]) and Vaibhav Gupta ([email protected]) to apply for the project.

Professor(s)

Alireza Karimi, Vishnu Varadan

Administration

Barbara Marie-Louise Frédérique Schenkel

Site

https://www.epfl.ch/labs/la/, https://www.epfl.ch/labs/ddmac/

The objective of this project is to set up an Active Suspension system test bench and apply data-driven robust control techniques on it. Data acquisition and controller implementation would be done using LabVIEW and NI DAQs. The designed controller would be robust to the parametric uncertainties in the system and will be verified and validated on the test bench.

Professor(s)

Alireza Karimi, Vishnu Varadan

Administration

Barbara Marie-Louise Frédérique Schenkel

Site

https://www.epfl.ch/labs/la/, https://www.epfl.ch/labs/ddmac/

This project aims to develop a novel control framework that bridges the gap between offline control and online control. The framework consists of two phases, i.e., experiment design, carried out without immediate performance considerations, and robust controller synthesis to minimize a performance index. As the quality of experimental data directly affects controller performance, the experiment is tailored to facilitate subsequent controller design. The project focuses on mathematically formalizing this pipeline and optimizing its components.

Professor(s)

Alireza Karimi, Zhaoming Qin

Administration

Barbara Marie-Louise Frédérique Schenkel

This project explores the implementation of a partial feedback linearization algorithm for the swing up controller of an inverted pendulum. Collocated and Non-collocated schemes will be explored. The algorithm will also be extended to the data-driven paradigm. The designed controller will be implemented on the hardware setup using MATLAB, Simulink.

Professor(s)

Alireza Karimi, Vishnu Varadan

Administration

Barbara Marie-Louise Frédérique Schenkel

Site

https://www.epfl.ch/labs/la/, https://www.epfl.ch/labs/ddmac/

Machine learning models are typically trained by minimizing an empirical loss function for the training dataset, which generalizes the model to other samples under the assumption that they follow the same distribution. In practice, however, data distributions can evolve in time, e.g., due to seasonality, user turnover, sensor drift, or other nonstationarities, leading to rapid degradation in out-of-sample performance.

Distributionally Robust Optimization (DRO) mitigates this risk by minimizing a worst-case expected loss over an ambiguity set of distributions centered at the empirical distribution. These sets are commonly defined via a statistical distance such as Wasserstein metric. An overview of DRO formulations for convex learning problems can be found in (Kuhn, 2019).

This project will aim to develop an online, time-adaptive DRO framework for machine learning problems under temporal distribution shifts. The approach (i) replaces the uniform empirical distribution with a recency-weighted empirical measure and (ii) adapts the robustness level over time based on data availability. We will formulate and solve classical convex machine learning problems in this framework, such as logistic regression and support vector machines.

Preliminaries:

– Strong mathematical background

– Convex optimization

– Python

References:

Kuhn, Daniel, et al. “Wasserstein distributionally robust optimization: Theory and applications in machine learning.” Operations research & management science in the age of analytics. Informs, 2019. 130-166.

Comment

Interested students should contact Jakob Nylöf ([email protected]).

Professor(s)

Giancarlo Ferrari Trecate, Lars Jakob Nylöf

Administration

Barbara Marie-Louise Frédérique Schenkel

Site

https://www.epfl.ch/labs/la/pi/