Student Projects

Projects are extracted from ISA database, some delays may occur.

For additional information and project status, please send an email directly to the project contact person/assistant.

Please note that the online project status (available/taken/etc) may not be accurate.

For Internship + PDM in indusry, please contact Dr. Alireza Karimi

Directives (2014) for projects at LA

Information to add/manage projects on ISA can be found here  (not official).

LA projects on ISA (Jones, Ferrari Trecate, Kamgarpour, Karimi, Salzmann)

ALT

Quick detection of the actuator subjective noise issues is a challenging task, especially if executed at the end of production line in noisy surrounding condition which prohibits actual direct measurement of acoustic noise emission. To tackle that problem, the indirect measurements of other variables such as current, voltage, and angular position of actuators obtained during the test are recorded. From these readings, one should be able to infer the state of the actuator.

In the project, we will look into developing a system Identification method and an AI module that would jointly enable the detection of faulty actuators.

Johnson Electric (JE) and EPFL are working together to develop a proof-of-concept tool that would solve the described problem.

Professor(s)
Giancarlo Ferrari Trecate
Administration
Nicole Anne Bouendin
External
Johnson Electric, Roman Klis, [email protected]
Site
https://la.epfl.ch/pi
ALT

Rapid detection of failures in manufactured parts can lead to substantial financial savings. This is especially true for high-volume production where the time slot dedicated to testing is highly limited. In state-of-the-art failure detection routines, a product undergoes a substantial testing procedure that cannot be shortened.

An option to introduce time-saving is to apply machine learning and system identification methods, which would shorten the failure detection time. In the thesis/project, the recorded measurements at end of line values such as current, and voltage angular position of actuators is going to be utilized to validate if shortening of testing time is possible.

In the project, we will look into developing a system Identification method and AI module which would jointly enable the detection of faulty actuators.

Johnson Electric (JE) and EPFL are working together to develop a proof-of-concept tool that would solve the described problem.

Professor(s)
Giancarlo Ferrari Trecate
Administration
Nicole Anne Bouendin
External
Johnson Electric, Roman Klis, [email protected]
Site
https://la.epfl.ch/pi
ALT

Rapid detection of failures in manufactured parts can lead to substantial financial savings. This is especially true for high-volume production where the time slot dedicated to testing is very limited. In state-of-the-art failure detection routines, a product process undergoes a substantial testing procedure that cannot be shortened.

An option to identify earlier the potential modes of failure is to leverage the model-based approach in which the identification of physically known systems such as electro-mechanical actuators can be carried out. Comparing the obtained results with reference ones could lead to the identification of failure modes, which then in turn can be utilized by unsupervised learning for data segmentation.

In the project, we will look into developing a system Identification method and AI module which would jointly enable the detection of faulty actuators.

Johnson Electric (JE) and EPFL are working together to develop a proof-of-concept tool that would solve the described problem.

Professor(s)
Giancarlo Ferrari Trecate
Administration
Nicole Anne Bouendin
External
Johnson Electric, Roman Klis, [email protected]
Site
https://la.epfl.ch/pi
ALT

The EPFL babyfoot is under continuous improvement.

While the babyfoot can easily intercept the ball, it has difficulty to capture it and make a pass. Juggling with the ball is another challenge. This project aims at improving the ball handling to permit capture and juggling, and then shoot toward opponents goal.

Students suggested improvements are also welcome.

Professor(s)
Christophe Salzmann (Laboratoire d’automatique 3)
Site
https://www.epfl.ch/labs/la/studentprojects/babyfoot/
ALT

Develop a control system for bumper cars where there are no collisions, no matter how hard kids try! Equip cars with local collision-avoidance controllers, smoothly overtaking control of the wheel/speed when cars are too close to each other.

Keywords: Automatic control, autonomous driving

Description

The project is developed within the NCCR Automation and is conceived for two master thesis.

Goal

PROJECT 1 – bumper car modeling modeling – customization of controllers for collision avoidance to the bumper car setting – comparison of alternative solutions in simulation

PROJECT 2 – design of miniature bumper cars (mechanical design + embedded controllers) – test using the tracking system available at the Automatic Control Lab at EPFL

Professor(s)
Giancarlo Ferrari Trecate
Administration
Nicole Anne Bouendin

Bayesian optimization, as an efficient black-box global optimization method (Frazier (2018)), has recently been applied to complex closed-loop controlled system optimization (e.g., Xu et al. (2021)). In particular, Bayesian optimization has been applied in robotics (e.g., optimizing trajectory time for quadrotor (Ryou et al. (2021)) and tuning the gaits for robot locomotion (Calandra et al. (2016))). This project aims to develop Bayesian optimization-based method and apply it to the automatic tuning problem arising from robotics. Proposed work may have a focus on either theory, algorithm or experiment, with the developed methodology deployed on Unitree A1 quadruped (https://www.unitree.com/products/a1/). Required skills include python or matlab programming and basic optimization. Previous experience with robotics, Robot Operating System (ROS), or Bayesian optimization is a plus, but not necessary.

1. Frazier, P. I. (2018). A tutorial on bayesian optimization. arXiv preprint arXiv:1807.02811

2. Xu, W., Jones, C. N., Svetozarevic, B., Laughman, C. R., and Chakrabarty, A. (2021). Vabo: Violation-aware bayesian optimization for closed-loop control performance optimization with unmodeled constraints. arXiv preprint arXiv:2110.07479

3. Ryou, G., Tal, E., and Karaman, S. (2021). Multi-fidelity black-box optimization for time-optimal quadrotor maneuvers. The International Journal of Robotics Research, page 02783649211033317

4. Calandra, R., Seyfarth, A., Peters, J., and Deisenroth, M. P. (2016). Bayesian optimization for learning gaits under uncertainty. Annals of

Mathematics and Artificial Intelligence, 76(1):5-23

Comment
Assistants: Shaohui Yang, Wenjie Xu
Professor(s)
Colin Neil Jones (Laboratoire d’automatique 3), Christophe Salzmann
Administration
Nicole Anne Bouendin
ALT

Commercial buildings consume around 40% of the world’s energy. Many research papers have explored different advanced methods to save energy. For example, model predictive control (MPC) does the job and guarantees occupancy comfort. However, there is still a large gap between the promising research results and real-world deployment. Thus, this project explores a more implement-friendly building control framework – data-driven control methods.

Current industrial practice is typically simple controllers such as bang-bang and PI controllers. To update these simple controllers to more advanced optimization/ learning-based control, such as predictive model control and reinforcement learning, additional investment in computation infrastructure is needed. One low-cost alternative solution is directly tuning those parameters of existing simple controllers to optimize the building performance. For example, Bayesian optimization has been demonstrated to be effective in tuning building controllers. This project could explore more experimental work on Bayesian optimization-based building controller tuning. A building testbed in the EPFL campus, Polydome (shown in the image below), is ready for actual experiments.

Requirements:

1. Proficient in Matlab or Python

2. Knowledge in optimal control or machine learning is preferred

3. Knowledge in Bayesian optimization or derivative-free optimization is a plus

Comment
Assistants: Jicheng Shi ([email protected]), Wenjie Xu ([email protected])
Professor(s)
Colin Neil Jones (Laboratoire d’automatique 3), Christophe Salzmann
Administration
Nicole Anne Bouendin
ALT

Novel-type high-precision optical instruments used for Earth observation missions require a very high pointing accuracy. Line-of-sight stability requirements constrain the admissible level of mechanical vibration on board spacecraft. Micro-disturbances are a phenomenon caused by satellite systems such as reaction wheels, thrusters, cryocoolers or solar array drive mechanisms. The term micro-disturbance refers to a mechanical vibration with low amplitude, typically occurring at frequencies from 1 Hz up to 1 kHz. These disturbances can substantially degrade the performance of sensitive payloads such as antennas, cameras or laser communication terminals. A hybrid active-passive micro-disturbance isolation platform has been developed at CSEM Neuchâtel in the frame of an ESA-funded PhD thesis. The modular demonstration platform consists of an adjustable number of passive dampers, a set of proof mass actuators creating a 6 DoF force tensor and an interface allowing to carry different types of payloads. A rapid prototyping platform is used for testing and optimization of different control algorithms.

The goal of this semester project is to perform a system identification of the platform at different operation points to synthesise a frequency-domain uncertainty model. Classical data-driven uncertainty identification approaches shall thereby be compared to a novel approach based on integral quadratic constraints (IQC). This method consists in finding the best linear approximation model combined with an elliptic uncertainty set. A comparison of the two methods in terms of performance and conservatism shall be made. The developed models shall be used in a later step for model-free control design (out of the scope of the semester project).

The experimental work shall be carried out in the Sensing&Control laboratory at CSEM Neuchâtel and the data processing and modelling can be performed at the Laboratoire d’Automatique at EPFL.

Professor(s)
Alireza Karimi, Elias Sebastian Klauser
Site
https://www.csem.ch/technical-focus/scientific-instruments

The large-scale integration of distributed power-electronic devices has rendered modern power systems difficult to be explicitly and accurately modeled through first principle or system identification. Meanwhile, the ubiquitous smart meters and smart sensors in power systems give us the access to a substantial amount of data. The behavioral approach, representing the system dynamics with trajectory data, lend itself to the analysis and control of complex power systems owing to non-parametric representation.

This project aims to design controller for complex microgrids using behavioral approach. First, the student can determine a specific task in microgrids, and start with the noise-free cases. Then, control methods against noisy measurement will be investigated. The performance of controllers will be validated and compared by simulation.

This project can either be a semester project or a master project.

Professor(s)
Alireza Karimi
Administration
Isabelle Stoudmann Schmutz
ALT

Motivation:

The EPFL Rocket Team aims for building thrust-vector-controlled rockets that can be reused by controlling them to land upright at a specific location. We develop algorithms that guide and control the rocket during the descent phase, approach, and touchdown.

Description:

In this project, we look into the planning of descent trajectories, i.e., how to get from the apogee (highest point of flight) to the desired landing location: When should the engine be re-ignited? What control algorithm stabilizes the rocket motion? Should there be active or passive aerodynamic control surfaces?

The ideal outcome of the project is a complete strategy for descent including the implementation of optimal guidance and control algorithms and testing in simulation. Results of this project might even affect the mechanical design of the next rocket.

Skills needed:

– Optimal Control (MPC)

– Knowledge about flight mechanics/reference frames is a plus

– Experience in C++ and coding projects is a plus

– Familiarity with ROS (robot operating system) is a plus

Comment
Please contact Johannes Waibel ([email protected]) if interested.
Professor(s)
Colin Neil Jones (Laboratoire d’automatique 3), Johannes Christian Karl Waibel
Administration
Nicole Anne Bouendin
ALT

Deep neural networks have proved successful in many application domains, such as image recognition, language comprehension, and sequential decision making. However, recently, neural networks are being employed for the control of non-linear dynamical systems.

In this project, our goal is to employ distributed neural networks for the control of interconnected systems while guaranteeing the closed-loop stability. Particularly, we will consider Hamiltonian deep neural networks for distributed control problems, for instance, formation, consensus, and coverage control with collision avoidance. Moreover, we will test our neural networks on the state-of-the-art platform called Robotarium that provides a remotely access to swarm robotics.

Comment
Assistants: Muhammad Zakwan ([email protected]) and Clara L. Galimberti ([email protected])
Professor(s)
Giancarlo Ferrari Trecate, Christophe Salzmann
Administration
Nicole Anne Bouendin
External
[email protected]
Site
https://www.epfl.ch/labs/la/pi/
ALT

In this project, we aim to develop a model-free optimization method as an alternative to Constrained Policy Optimization for Reinforcement Learning tasks. The application problem we will address is collision-free multi-vehicle trajectory planning.

The main idea of our approach is to parameterize the possible policies and optimize over the parameters in order to reduce the cost. Specifically, we use samples to form local approximations of the cost and the safety constraints.

The challenge of this method mainly lies in the sampling of the cost and the constraints, which are always the expectations of random variables. The existing methods to form good approximations of the expectations usually require a lot of samples while the sampling is always expensive. One possible approach is to model the samples by machine learning.

The experiments are implemented on Robotarium and Safety Gym (OpenAI). The student will first construct a barrier function and a control strategy to avoid vehicle collision. Then improvement over this strategy is carried out through zeroth-order optimization to reduce the cost while maintaining safety.

This project can either be a semester project or a master project

Prerequisite: solid programming skills, decent ability to understand literature and express a complex concept

Contact: Baiwei Guo, [email protected]

Professor(s)
Giancarlo Ferrari Trecate, Christophe Salzmann
Administration
Nicole Anne Bouendin
Site
https://www.epfl.ch/labs/la/pi/
ALT

Photovoltaic (PV) cells installed in residential buildings are a promising renewable energy resource imposing minimal damage to the environment. However, the high variability of the generated power with atmospheric conditions impedes the full integration of PV cells. Machine learning algorithms were employed to predict PV power generation at short time intervals from meteorological features and power generation at previous time steps to alleviate this condition.

While there has been a vast amount of work on predicting PV generation employing various machine learning algorithms, the proposed approaches suffer from requiring long data collection periods before the models can be used and are sensitive to the change of cells’ characteristics due to aging or technological renovations. Additionally, most methods rely on solar irradiance as a strong predictor of solar power output, neglecting the tedious procedure of irradiation measurement and the high error of its approximation. Therefore, it is desirable to develop an algorithm for predicting PV generation at a high frequency from features measured by households that can be trained on a relatively small training dataset.

A potential source of acquiring more information for training a prediction model at a specific residence is through consulting with other households. A naive approach is to augment the training dataset available to different households and train a model on it, which is not conducive due to a variety of cell technologies, installation setups, and geographical locations among households. Still, since the underlying physical phenomenon for PV generation is the same, it is insightful to develop a scheme in which households collaboratively train their prediction models while communicating relevant information. Federated Learning is a well-tailored framework for establishing collaboration among the participating data holders, in our case, the households, to extract common information. We will employ Federated Learning to train an XGBoost model, which has been found prosperous in multiple time-series prediction tasks.

In this project, the student is asked to apply state-of-the-art federated-learning algorithms to train an XGBoost model for predicting PV power generation. This proposal is an applied project and requires strong programming skills, plus a good knowledge of machine learning.

Comment
Contact person: [email protected]
Professor(s)
Giancarlo Ferrari Trecate, Mahrokh Ghoddousiboroujeni
Administration
Nicole Anne Bouendin
ALT

In this project, the objective is first identify a model for a 3-Degree-Of-Freedom (3DOF) hover based on conventional techniques in system identification. To do so, first, a code to run the system using LABVIEW should be written and deployed on MyRIO.

Moreover, all the electrical connections should also become compatible for running the system.

The code for system identification should be flexible to select different excitation such as PRBS, white noise, single tone sinusoidal and sin-sweep. The second part of the project includes designing a controller using the methods of advanced control such as H-infinity and data-driven method and then applying on the device using an appropriate labview code and validate the performance.

Comment
Assistant: Vaibhav Gupta
Professor(s)
Alireza Karimi, Vaibhav Gupta
Administration
Isabelle Stoudmann Schmutz
Site
la.epfl.ch
ALT

In this project, the objective is first identify a model for a ball and plate system based on conventional techniques in system identification. To do so, first, a code to run the system using LABVIEW should be written and deployed on MyRIO. Moreover, all the electrical connections should also become compatible for running the system. The code for system identification should be flexible to select different excitation such as PRBS, white noise, single tone sinusoidal and sin-sweep. The second part of the project includes designing a controller using the methods of advanced control such as H-infinity and data-driven method and then applying on the device using an appropriate LABVIEW code and validate the performance.

Comment
Assistant: Vaibhav Gupta
Professor(s)
Alireza Karimi, Vaibhav Gupta
Administration
Isabelle Stoudmann Schmutz
Site
ddmac.epfl.ch, la.epfl.ch
ALT

Motivation:

In the previous semesters, we build a lightweight hovercraft platform capable of hovering on an air-hockey table. The long-term goal of the project would be the autonomous play against humans, or having multiple hovercrafts playing air hockey autonomously against each other.

Description:

In this project, we look into the identification and control of the hovercraft. For this, we aim to utilize novel machine learning techniques to identify the stabilizable dynamics of the system from data. Using this model, we intend to use advanced control algorithms (e.g. time-optimal MPC or reinforcement learning) to steer the hovercraft at its physical limits. For this, you will build upon our existing ROS2 software stack, integrate your work, and assess the performance of your algorithms both in simulation and hardware.

Skills needed:

– Excellent knowledge of C++ and Python

– Basic Machine Learning knowledge

– Knowledge about Optimal Control (MPC) is a plus

– Familiarity with ROS/ROS2 is a plus

If interested, please contact the responsible assistant directly ([email protected]) and attach your transcripts and CV.

Professor(s)
Colin Neil Jones (Laboratoire d’automatique 3), Roland Schwan
Administration
Nicole Anne Bouendin
ALT

Many optimal controller synthesis procedures are based on parametrizing all stabilizing controllers, and the responses that they achieve, over which relevant performance measures can be easily optimized. For the linear time-invariant case, the Youla parametrization allows us to directly design the system response rather than synthesizing the controller itself. For the more general case, a nonlinear Youla-like parametrization can also be adopted. Moreover, there is a relationship between this parameter and the corresponding closed loop maps [1].

An obstacle to directly applying these results in practice is that the operators are infinite-dimensional. Hence, a finite-dimensional approximation is necessary at the implementation step. For the nonlinear case, recurrent equilibrium networks (RENs) [2] have shown to be flexible models for parametrizing nonlinear stable operators. However, further models could be used such as port-Hamiltonian networks [3] or dynamic recurrent neural networks [4].

The first goal of the thesis is to develop variants of RENs to broaden the parametrization of stable nonlinear operators. Besides, many real-world applications involve large-scale systems where local controllers only have access to limited observations. Thus, the second goal of the thesis is to adapt the control policies for the distributed scenario. Relevant large-scale applications will be considered to test the effectiveness of the developed algorithms.

[1] Furieri, L., Galimberti, C., and Ferrari-Trecate, G. “Neural System Level Synthesis: Learning over All Stabilizing Policies for Nonlinear Systems.” arXiv preprint arXiv:2203.11812, 2022.

[2] Revay, M., Wang, R., & Manchester, I. R. “Recurrent equilibrium networks: Flexible dynamic models with guaranteed stability and robustness.” arXiv preprint arXiv:2104.05942, 2021.

[3] Furieri, L., Galimberti, C. L., Zakwan, M., & Ferrari-Trecate, G. “Distributed neural network control with dependability guarantees: a compositional port-Hamiltonian approach.” In Learning for Dynamics and Control Conference (pp. 571-583). PMLR, 2022.

[4] Kwang-Ki K. Kim, Ernesto Ríos Patrón, Richard D. Braatz, “Standard representation and unified stability analysis for dynamic artificial neural network models,” Neural Networks, Volume 98, 2018.

Comment
Contact: Clara Galimberti ([email protected])
Professor(s)
Giancarlo Ferrari Trecate, Clara Lucía Galimberti
Administration
Nicole Anne Bouendin
Site
https://www.epfl.ch/labs/la/pi/
ALT

Robotic quadrupeds have received an increasing amount of attention in recent years. Academically speaking, their complex nonlinear and hybrid dynamical nature as well as contact-rich behaviors lead to exciting challenges to be solved to control such platform properly. Nonlinear Model Predictive Control (NMPC) is a representative of the state-of-the-art model base/free methods that are capable of performing highly agile and dynamic legged locomotion. The robotics community focuses more on model simplification and uncertainty handling for real world deployment, while the control community pays more attention on the inherent nature of the platform to exploit structure. Both are aiming at developing numerical methods efficient enough for online usage. However, basic properties of MPC such as recursive feasibility and stability has long been neglected.

In this project, the student will first study the current features of NMPC applied on quadruped, followed by designing stage-wise and terminal cost/constraints specifically for quadruped such that the nice properties on MPC can be retained. The final stage will be developing C++ code and doing real-world experiment to validate the proposed strategy. Depending on the student’s background and interest, the project may be designed at a suitable balance between theory and application. The content of the project is also subject to change, with the negotiation between the student, the supervising PhD student and Professor, but the main direction will be model-based planning and control for quadruped locomotion.

Comment
Course requirements: Model Predictive Control, Optimization

Software requirements: C++, Python, ROS

Doctoral assistant: Shaohui Yang

Professor(s)
Colin Neil Jones (Laboratoire d’automatique 3), Christophe Salzmann
Administration
Nicole Anne Bouendin
Site
https://www.epfl.ch/labs/la/pi/
ALT

Motivation:

The EPFL Rocket Team aims for building thrust-vector-controlled rockets that can be reused by controlling them to land upright at a specific location. We develop algorithms that guide and control the rocket during the descent phase, approach, and touchdown.

Description:

In this project, we would like to develop algorithms for safely adapting drone and rocket controllers online, i.e., while they are flying. This may involve classical & optimal control methods, or machine learning. Another approach is to conduct online identification, i.e., improve a model of the system, and use the identified model to derive controllers.

The ideal outcome of the project is a reliable tuning procedure that is thorougly tested in simulation and experimentally on a rocket-like drone. The resulting algorithm could later be applied to a 50kg hopper with a rocket engine.

Skills needed:

– Classical and/or Optimal Control (MPC)

– Experience with Machine Learning methods is a plus

– Experience in C++ and coding projects is a plus

– Familiarity with ROS (robot operating system) is a plus

Comment
Please contact Johannes Waibel ([email protected]) if interested.
Professor(s)
Colin Neil Jones (Laboratoire d’automatique 3), Johannes Christian Karl Waibel
Administration
Nicole Anne Bouendin
ALT

Predicting the electricity consumption of residential buildings is fundamental to scheduling shiftable loads and flattening the demand curve. While there has been a vast amount of work on predicting electricity consumption employing various machine learning algorithms, the proposed approaches suffer from requiring long data collection periods before the models can be used. Therefore, it is desirable to develop an algorithm for predicting households’ electricity consumption that can be trained on a relatively small training dataset.

A potential source of acquiring more information for training a prediction model at a specific residence is through consulting with other households. A naive approach is to augment the training dataset available to different households and train a model on it, which is not conducive due to the variation in the households’ consumption habits. Still, since the underlying physical patterns for energy consumption are the same, it is insightful to develop a scheme in which households collaboratively train their prediction models while communicating relevant information.

Personalized Federated Learning is a well-tailored framework for establishing collaboration among the participating data holders, in our case, the households, to extract common information while adapting to individual datasets. The main idea is to capture shared patterns with a generic model, called the global model, and adapt it to individual datasets to train personalized models accounting for personal differences.

In this project, the student is asked to apply state-of-the-art federated-learning algorithms to the problem of high-frequency prediction of residential electricity consumption. This proposal is an applied project and requires strong programming skills, plus a good knowledge of machine learning.

Comment
Contact person: [email protected]
Professor(s)
Giancarlo Ferrari Trecate, Mahrokh Ghoddousiboroujeni
Administration
Nicole Anne Bouendin
ALT

A reliable forecast of the power consumption and generation of prosumer households would provide a great improvement in the energy ressource management on distribution grids. Unfortunately, this task is very challenging due to the unpredictability of both the weather and human behavior. This project aims at comparing the performance of a wide variety of machine learning methods, and making use of as much data as possible, in order to maximize the accuracy of the predictions.

A real set of timeseries collected on a building equipped with solarpanels, heat pumps, EV chargers, and a meteo station, as well as historical predictions from Meteoswiss will be used to test and train the models. The student should identify the challenges precisely and provide an analysis of the tools available to tackle them.

Skills required:

-Good knowledge of State of the art ML techniques (regressions, trees, NNs)

-Excellent python skills

-Experience with large datasets

For registration or for more information, send a mail to [email protected] or [email protected] with your CV and transcript. Applications will be collected until January 30th and a student will be selected then.

Comment
Assistants: Jean-Sébastien Brouillon, Muhammad Zakwan
Professor(s)
Giancarlo Ferrari Trecate, Christophe Salzmann
Administration
Nicole Anne Bouendin
Site
https://la.epfl.ch/pi
ALT

In recent years, wind farms have covered increasingly larger shares of green energy production. To support this rapid development of sustainable energy systems, it is crucial to design optimal control schemes that allow to dynamically dispatch the wind farm power demand to individual wind turbines while minimizing mechanical stress [1].

As the dynamics of each turbine are affected by turbulent winds, however, optimally controlling these systems proves challenging. In this thesis, we propose to design controllers that can adapt to realistic wind profiles by minimizing regret, that is, the loss relative to an ideal clairvoyant policy that has noncausal access to past, present, and future disturbances [2]. The performance of existing and newly developed algorithms will be assessed by means of high-fidelity wind farm simulator models.

Knowledge about linear system theory and programming experience with Matlab or Python is required.

[1] Riverso, S., Mancini, S., Sarzo, F., & Ferrari-Trecate, G. “Model predictive controllers for reduction of mechanical fatigue in wind farms.” IEEE Transactions on Control Systems Technology, 25(2), pp. 535-549, 2016.

[2] Martin, A., Furieri, L., Dörfler, F., Lygeros, J., & Ferrari-Trecate, G. “Safe control with minimal regret.” In Learning for Dynamics and Control Conference (pp. 726-738). PMLR, 2022.

Comment
Contact: [email protected]
Professor(s)
Giancarlo Ferrari Trecate, Andrea Martin
Administration
Nicole Anne Bouendin
ALT

Motivation:

The EPFL Rocket Team aims for building thrust-vector-controlled rockets that can be reused by controlling them to land upright at a specific location. We develop algorithms that guide and control the rocket during the descent phase, approach, and touchdown.

Description:

In this project, we would like to develop robust control algorithms for an autonomous rocket touchdown maneuver. Challenges are short time disturbances, ground effect, and gear-ground interaction.

The ideal outcome of the project is a reliable controller implementation and thorough testing both in simulation and experimentally on a rocket-like drone. The resulting controller may be used in a 50kg hopper with a rocket engine.

Skills needed:

– Classical and/or Optimal Control (MPC)

– Experience in C++ and coding projects is a plus

– Familiarity with ROS (robot operating system) is a plus

– Knowledge about flight mechanics/reference frames is a plus

Comment
Please contact Johannes Waibel ([email protected]) if interested.
Professor(s)
Colin Neil Jones (Laboratoire d’automatique 3), Johannes Christian Karl Waibel
Administration
Nicole Anne Bouendin

Policy Optimization (PO) is a popular reinforcement learning method. Given a specific task, the main idea of PO is to parameterize the possible policies and optimize over the parameters in order to reduce the cost. In this project, we focus on safe model-free PO, which uses samples to form local approximations of the cost and the safety constraints.

The challenge of this method mainly lies in the sampling of the cost and the constraints, which are always the expectations of random variables. The existing methods to form good approximations of the expectations usually requires a lot of samples while the sampling is always expensive.

To solve this issue, the student can either use the techniques from stochastic programming to reduce the samples required or improve upon the efficiency of deriving a single sample.

This project can either be a semester project or a master project

Prerequisite: solid programming skills, decent ability to understand literature and express a complex concept

Contact: Baiwei Guo, [email protected]

Professor(s)
Giancarlo Ferrari Trecate, Christophe Salzmann
Administration
Nicole Anne Bouendin
Site
https://www.epfl.ch/labs/la/pi/
ALT

Unsupervised learning has recently demonstrated a large potential for application in the quality control field. This is especially true for high-volume production where obtaining laboratory-tested samples can be challenging.

The application of unsupervised learning based on time series data could lead to discover failure modes that were not known to the quality department earlier. In the thesis/project the recorded values at end of line such as current, and voltage angular position of actuators are going to be utilized to validate if time-series data obtained from quality control can enable detection/separation of failure modes.

In the project, we will look into developing a system Identification method and AI module which would jointly enable the detection of faulty actuators.

Johnson Electric (JE) and EPFL are working together to develop a proof-of-concept tool that would solve the described problem.

Professor(s)
Giancarlo Ferrari Trecate
Administration
Nicole Anne Bouendin
External
Johnson Electric, Roman Klis, [email protected]
Site
https://la.epfl.ch/pi

Industrial projects (not yet on ISA)

To address climate change properly and accelerate energy transition from fossil fuels to renewable energy, new energy generation technologies are required. They have to be reliable, efficient, scalable and provide sufficient low-cost energy to satisfy current and future demands. Airborne Wind Energy (AWE) is an emerging technology that involves autonomous tethered aircraft to produce electricity from high altitude winds. 
Skypull can offer a solution and proposes an innovative approach, based on a new revolutionary type of flight device, aimed at maximizing the lift, with minimum resistance increase and allowing autonomous operation: a “box” drone with proles multi-element aerodynamics, characterized by high aerodynamic performance, intrinsic structural strength, low weight and low production costs. 
In order to keep the drone operational over prolonged periods of time it is critical to recharge the onboard electronics during operation. Skypull developed the technology of regenerative differential braking, which allows the controlled braking of the aircraft while simultaneously regenerating energy. 

Goal 

In order to better understand the process of differential braking our drone model needs to be extended to accommodate for the braking process. The models focus will be on the drivetrain (motor controller, motor and propeller), but should be integrated in the drone model. 

Based on this model a set of control laws will be developed to enable the drone to brake, turn and regenerate energy in a stable and robust way.

Responsibility 

* Literatue/Skypull technology review 
* Installation of a test bench 
* Implementation of the drive train model and integration into drone model 
* Derivation of control strategy for differential braking 
* Testing on Skypull hardware

Requirements 

* Good understanding of Simulink / MATLAB 
* Good fundamental knowledge in control system design and modelling 
* Basic knowledge of aerodynamics and aircraft actuators 
* Basic knowledge of motor controllers 
* Ability to work independently, proactive personality and highly motivated 

Professor:Colin Jones

Type of Project: Master in Industry

Industrial contacts 

Andrea Pedrioli Simulation Engineer, [email protected]

Aldo Cattano Chief Technology Officer, [email protected] 
Skypull SA, via alla Stampa 49, 6967 Lugano

Literature 

* Loyd M. L., “Crosswind kite power”, Journal of Energy 4(3) 106-111 (1980). 
* Kathryn E. Johnson , “Adaptive Torque Control of Variable Speed Wind Turbines”

The energy systems group at CSEMis developing innovative technologies at the intersection of power and energy, control, and data science. In the march towards a fully- decarbonised energy system, we are developing control solutions to make buildings and grids more energy-efficient and more flexible. The current state of the art is model- predictive control to optimize the operating conditions of distributed energy resources. However its practical deployment is hindered by the time and effort required to build models of these resources, and by the inability of most models to reflect the impact of occupants on the energy performance of buildings.

In this context, we are offering an internship on the development of artificial-intelligence algorithms. The objective of the project is to build on state-of-the-art algorithms in deep reinforcement learning to create robust and optimal control strategies. Such algorithms include deep Q-networks (DQN, first demonstrated by Deepmindin 2013 to play Atari games), deep deterministic policy gradient (DDPG, first introduced by Google Deepmindin 2015 for simple motion control tasks), and proximal policy approximation (PPO, introduced by Open AI in 2017 for humanoid running and steering).

The physical nature of the control problems in energy management raises some specific challenges for these algorithms, especially in terms of convergence. The project will therefore combine theoretical aspects, with an opportunity to improve some of the most advanced general-purpose artificial intelligence algorithms, and practical aspects involving the rapid prototyping of software architectures and the implementation of algorithms.

While the objective of the work is clearly industrial, its nature means there will be scope to present the results in scientific journals or at conferences. The student will be part of a team of twelve experienced engineers and researchers.

Professor:Colin Jones

Type of Project: Master in Industry

Contacts: Pierre-Jean Alet, Baptiste Schubnel

Autonomous driving technologies that aim for more safety, comfort, and environmentally friendly vehicles, have been growing rapidly in automotive industry recently. These systems go beyond vehicle-centric assistant applications by integrating traffic and environment information. One of the main challenges in autonomous driving development is to design and then validate the planning and control algorithms in a closed-loop fashion, where both vehicle dynamics characteristics and a wide variety of traffic scenarios are taken into account. The designs should also guarantee optimal performance toward precise tracking, and time/fuel optimality with respect to both vehicle and traffic constraints (i.e. avoid pedestrians and other cars).

One of the main innovation objectives of Siemens PLM is to accelerate the research and development activities of ADAS (advanced driver assistance systems) and autonomous driving technologies. We are continuously looking for outstanding students who are eager to do their Master thesis or internship on a challenging research project in a highly dynamic and international research environment. We have a variety of possible projects available that cover different aspects of planning and control algorithms, ranging from very theoretical to practical. The company also provides various tools to support the research activities, for example, Imagine.Lab Amesim for vehicle dynamics modelling, PreScan for sensor (camera, lidar, radar,…) and traffic environment modelling, and a miniature race car setup for embedded control implementation.

Examples of possible (but not limited to) topics: optimal control and learning control for an ADAS application (i.e. roundabout crossing, valet parking), fast embedded optimal control algorithm, optimal path/motion planning, simultaneous localization and mapping (SLAM) in virtual environment, vehicle dynamics, traffic planning and verification…

Background: control systems, robotics, computer science, or mathematics, familiar with programming. Experience with optimization, vehicle dynamics, ROS or autonomous vehicles is a plus.

Professor: Colin Jones

Type of project: Master

Contact: Colin Jones or Dr. Son Tong

Image result for siemens logo The work will be performed in collaboration with Siemens PLM Software in Leuven, Belgium.

Houses are commonly equipped with PV panels, electric batteries, variable power heat pump and electric vehicles. This means single homes are becoming complex energy hubs which require an efficient energy management system (EMS) that can be configured in a simple way while satisfying complex needs such as improving self-consumption and reducing cost.

The energy systems group at CSEM is actively developing a next-generation energy management system that is based on Model Predictive Control and which can be easily reconfigured through an automated controller synthesis process. A first internship successfully contributed to this project with work focused on forecasting. In this context, we are offering an internship to improve the operation of this EMS by extending its functionalities, focusing on two aspects:

  • The refinement of the predictive control algorithm with specification of complex objective such as peak shaving in the EMS algorithm, and the scale-up to multi-home systems.
  • The buildup of a simulation environment that allows to testbench the controller. The simulation will rely on the versatile physical simulation environment Modelica, for which a collection of house models has already been developed. It will be used to assess the controller performance over long periods of time.

While the objective of the work is clearly industrial, there will be scope to present the results in scientific journals or at conferences. The student will be part of a team of twelve experienced engineers and researchers, and have the opportunity to use real data from test sites and face actual practical challenges of software deployment for energy applications.

Requirements:

  • A strong interest for energy systems and energy management problematics.
  • Good programming experience, preferably with exposure to Python; and an interest in good coding practices (object-oriented programming, unit testing, version tracking with Git).
  • Familiarity with MPC or numerical optimization preferred, knowledge of Kalman filtering and/or machine learning a plus.

Type of project: Master in industry or internship, minimum 4 months.

Location: Neuchatel

Contact: Tomasz Gorecki, [email protected]

Professor: Colin Jones, [email protected]


Past projects (not yet ready)

Projects 2018

Projects 2017

Projects 2016

Projects 2015

Projects 2014

Projects 2013

Projects 2012