Open Master's Thesis Positions

This research is a parallel investigation in collaboration with a researcher at the Robotic Fabrication Lab (RFL) at the D-ARCH department, where the paste iterations can be subjected to the large-scale printing and sintering experiments during development. The student will carry out the following tasks: • Characterisation of clay composition • Formulation and iterative testing of candidate organic pastes • Rheological characterisation to assess printability • Extrusion/casting and drying experiments to characterise shrinkage of the developed paste The candidate should meet the following criteria: • Interest in additive manufacturing technologies • Ability to independently plan and carry out hands-on material experiments • Knowledge of material characterisation techniques • Background in soft materials or a related field is an advantage

The main aim of this project is to develop a printable paste with an organic composition that is compatible with clay and reduces the overall shrinkage of the printed component during drying and sintering. The paste should be developed according to the following criteria: • Sufficient viscosity for printability and stable extrusion at 2 - 8 mm • Sufficient stability for layering • Low shrinkage during the drying process • Fully and safely combustible composition

Please send your CV to lim@arch.ethz.ch. This project can be undertaken as an open project, a semester project or a master's thesis, and the scope will be adjusted accordingly. Feel free to send an email if you have any questions. More details will be given about the project during the initial discussions.

In this project, we aim to learn stable control policies by exploiting information that trajectory optimization already provides but usually discards. Under standard assumptions, the optimal cost of the trajectory optimization problem, viewed as a function of the initial state, acts as a Lyapunov function: it decreases along the trajectories of the optimal controller. Moreover, its sensitivity with respect to the initial state can be obtained at almost no extra cost from the Lagrange multipliers at the solution, and it indicates which directions of motion decrease the value. Rather than training the policy to merely reproduce the expert's actions, this approach penalizes violations of the descent condition for the expert's value function (or an approximation thereof), encouraging Lyapunov stability as part of the learning objective. We therefore structure the method in two stages operating at different levels of computational cost: 1. **Offline expert and value-function learning.** We solve trajectory optimization problems offline from many initial states, recording the optimal inputs, the optimal cost, and its derivatives with respect to the initial state. From these samples, a smooth global approximation of the value function is constructed (e.g., by blending local first-order models or by fitting a gradient-supervised neural network), so that the Lyapunov candidate can be evaluated and differentiated at states that were never sampled. 2. **Online learned policy.** A parameterized policy, either a neural network or a differentiable MPC with learnable cost weights, terminal ingredients, and constraint-tightening margins, is trained to run in real time. Its parameters are adjusted through a gradient-based procedure that combines expert imitation with a penalty for violations of the Lyapunov descent condition along the system dynamics. If applicable and theoretically possible, parameters of the learned policy may also be adjusted online. We will evaluate the proposed algorithm in simulation on benchmark nonlinear systems with state and input constraints. We will assess the approach's performance by comparing closed-loop stability, region of attraction, constraint satisfaction, and closed-loop cost against the expert controller and against policies trained by pure imitation, using the same policy class and model.

The central contribution of this project is to turn trajectory optimization data into stability-aware learned controllers: by penalizing violations of the descent condition for the expert's value function or an approximation thereof during training, we aim to obtain policies that retain most of the expert's performance while providing verifiable closed-loop stability, without requiring additional online computations for the descent objective when the policy parameters are fixed. - Implement an offline data-generation pipeline that solves the trajectory optimization problem to record optimal inputs and costs, and extracts the value function derivatives using the optimal Lagrange multipliers. - Develop a smooth approximation of the value function and a training scheme that combines imitation and Lyapunov descent, and apply it to neural-network or differentiable-MPC policies. Depending on the project scope, compare both policy classes. - Validate the learned policies in closed-loop simulation, comparing descent-based training against imitation-only baselines, and study how the accuracy of the value-function approximation affects the stability of the resulting closed loop.

Please email your resume/CV (including lists of relevant publications/projects) and transcript of records in PDF format to johnemus(at)ethz.ch and rzuliani(at)ethz.ch.

Industrial contact-rich manipulation tasks, such as polishing and assembly, are inherently geometric tasks where directional information—such as the insertion axis and local surface normals—is critical. Current policy learning methods [1] leverage vision as the main modality, which requires extensive data to learn this directional information. In contrast, point clouds are better suited for geometric reasoning and can be explicitly augmented with directional information. However, existing point cloud encoders [2] are rarely evaluated on industrial tasks, and they do not incorporate task-relevant directions [3]. This project focuses on developing a point cloud encoder for physical AI applications and combining it with other modalities from TACTO1 (see inline Figure) to achieve superior generalization across novel environments and target objects. Conducted in collaboration with Bota Systems AG, this project offers direct exposure to real-world physical AI applications within the modern robotics industry.

The goal of this project is to develop a point cloud encoder for learning force-aware manipulation policies with superior geometric generalization compared to RGB-only policies. The encoder processes spatial points augmented with pre-computed local and global directions, such as surface normals, principal curvatures, and the nominal insertion axis. The output of this encoder is then fused with latent embeddings from a wrist force/torque sensor and the end-effector pose. By encoding explicit geometric priors rather than relying on raw RGB pixels, the primary objective is to improve cross-environment generalization and policy transfer relative to standard vision-based architectures. The main goals of the project are as follows: - Integrating depth/point cloud streams into the TACTO setup and developing automatic extrinsic calibration routines for wrist and workspace cameras; - Developing a specialized point cloud encoder capable of processing directional vector inputs; - Combining the point cloud encoder’s output with latent force and proprioception embeddings to condition the diffusion policy; - Evaluating the system experimentally and benchmarking its cross-environment and cross- object transfer capabilities against standard vision and baseline point cloud encoders. This work combines elements from robotics, 3D computer vision, and machine learning, with a strong focus on real-world hardware implementation.

Please email your resume/CV (including relevant projects) and transcript of records in PDF format to Cem Bilaloglu (cem.bilaloglu@inspire.ch)

Industrial contact-rich manipulation tasks, such as polishing and insertion, involve complex transitions across multiple states—including free-space motion, impact, sliding, sticking, and separation. Reliably registering and anticipating these events is critical for executing proactive control actions. Different sensor modalities offer complementary advantages for cap- turing these state changes: while wrist force-torque sensors register interactions reactively after contact occurs, visual data and point clouds provide proactive geometric cues. For instance, as a robot approaches a workpiece, spatial data allows the system to anticipate the shift from free space to contact before forces peak. This project focuses on developing a learning-based framework for physical AI that leverages rich multi-modal sensor streams from TACTO1 (see inline Figure) to anticipate contact transitions ahead of time. Conducted in collaboration with Bota Systems AG, this project offers direct exposure to real-world physical AI applications within the modern robotics industry.

The goal of this project is to develop a multimodal transformer that recognizes early precursors to transitions between contact states, such as sticking, sliding, and separation, enabling proactive robotic control. The model will be trained and evaluated on multimodal data from both simulation and the real world. Inputs will include gravity-compensated external wrenches, RGB images, point clouds, and estimated robot motion. Training labels will be generated automatically through physics-informed weak supervision, combining first-principles contact mechanics with the existing force, motion, and interaction observers. These mechanics-based detectors provide causal estimates of the current contact state and therefore serve as physically grounded training labelers. The central research objective is to determine whether a multimodal transformer can exploit the richer multimodal sensory history to identify precursors to contact-state transitions before those transitions can be confirmed by the causal mechanics-based detectors used to generate its training labels. The resulting predictor should generalize across simulated and real-world interaction data and provide sufficiently early and reliable transition estimates to support proactive adaptation of the robot controller. The main goals of the project are as follows: - Collecting multi-modal real-world (TACTO) and simulated data streams (Isaac Lab); - Formulating contact-state labeling via first-principles contact mechanics and force-motion ob- servers; - Training a multimodal transformer to predict upcoming contact transitions; - Evaluating the system experimentally and benchmarking its anticipation latency against pure mechanics-based detectors. This work combines elements from machine learning, robotics, control theory, and contact mechanics with a strong focus on real-world hardware implementation.

Please email your resume/CV (including relevant projects) and transcript of records in PDF format to Cem Bilaloglu (cem.bilaloglu@inspire.ch).

This project is set to limited visibility by its publisher. To see the project description you need to log in at SiROP. Please follow these instructions: - Click link "Open this project..." below. - Log in to SiROP using your university login or create an account to see the details. If your affiliation is not created automatically, please follow these instructions: http://bit.ly/sirop-affiliate

At the core of HVAC operation is the process of heat exchange, where a controllable primary fluid exchanges heat with a secondary fluid. The objective is to control the outlet state of the secondary fluid in order to satisfy a set of indoor environmental requirements in terms of temperature, indoor air quality (CO2), and occupant comfort. While the thermodynamic principles describing heat exchange are well understood, their inherent complexity, non-linear behavior, and internal system constraints arising from the real-world operation render the control task notoriously difficult. Following ideas from recent literature the task of how to create data-driven parametrized models from data that ensure the satisfaction of a set of priors will be pursued. Extending the current literature, which mainly focuses on autonomous dynamics, studying the influence of the control action. Following the core idea of Meta-Learning, strategies for updating the parametric model based on the controller outcome will be analyzed and extended, ensuring the adapted model consistently preserves priors. The algorithms developed will be validated against high-fidelity simulation environment provided by an industrial-partner. Extension to real test-bench data is also possible.

The goal of the project is to develop online data-based identification and control algorithm which adapt to unknown scenarios and work in cross-domain solutions. The algorithms will be validate on detailed simulation environments coming from industrial partners and possibly on real test-bench data. The project can be divided in the following six main objectives: 1. Literature review on HVAC systems and thermodynamical priors; review on physics informed ML algorithms. 2. Define a parameter-based model ( N.N, Kernel-based, etc..) to learn the system dynamics. Formulate strategies for incorporating priors ( e.g., projection layers, physics-based structures such as PINNs, Hamiltonian-Systems, etc...) 3. Test the obtained model against a detailed simulation environment from industrial partner. 4. Formulate an optimal constrained control problem for temperature tracking ( possibly also additional references ) incorporating the data-based model. 5. Formulate a prior-based meta-learning update of the system's parameters based on the outcome of the controller. 6. Validation of the methods in a high-fidelity simulation environment. 7. (optional) Extend the results on test-bench data.

To apply send an email with your resume/CV, including lists of relevant publications and projects, and transcript of records in PDF format to Alessandro Cupo at calessandr@control.ee.ethz.ch and Aren Karapetyan at aren.karapetyan@belimo.ch

Please have a look at the attached document.

Please have a look at the attached document.

Please email your CV (including lists of projects and optionally any document helpful to evaluate your background) and current grade transcript in PDF to: schandraseka@ethz.ch. In addition, if you have any questions on the project, please do not hesitate to contact me.

Model Predictive Control (MPC) is an industry-standard technique capable of achieving incredible performance while respecting hard safety constraints. However, its success relies heavily on two things: a meticulously tuned cost function and a highly accurate system model. When these elements are unavailable or too expensive to engineer, researchers turn to learning-based MPC. By combining online learning with traditional control, the controller can adapt on the fly to its environment, for example adjusting for mechanical wear, changing weather, or shifting payloads. However, while promising, learning-based MPCs are rarely adopted in practice because they can be unreliable. Without strict guardrails, these adaptive mechanisms can behave unpredictably and destabilize the system. This project aims to close this gap: you will develop learning-based MPC schemes that adapt dynamically without compromising safety, guaranteeing that the system won't break while it learns.

To achieve safe learning, this project leverages a modern, cutting-edge framework: _differentiable MPC_ [1]. Traditionally, an optimization solver is treated as a black box. Differentiable MPC opens this box, allowing us to compute the exact derivatives of the MPC solution map (how the current state maps to the applied input). By extracting these derivatives, we can tune and design the MPC scheme, modifying its cost function or constraints, using the same powerful, gradient-based optimization techniques that drive modern machine learning. Ultimately, this allows us to enforce rigorous safety guarantees during the learning process. The project will start with a review of several learning-based MPC approaches, followed by a study on how these can be enhanced through differentiable optimization. The simulations will be implemented on a readily available software base developed in our lab. Given the cutting-edge nature of differentiable control, you will have the opportunity to work towards a publication on this high-impact topic. **Project Tasks** - Familiarize with the code base (1 month). - Implement learning-based MPC schemes (1 months). - Enhance learning-based MPC with differentiable optimization (2 months). - Implement realistic simulations (1 month). - Document the results and write a report (1 month). **References** [1] Zuliani, Riccardo, Efe C. Balta, and John Lygeros. "Closed-loop performance optimization of model predictive control with robustness guarantees." European Journal of Control (2025): 101319.

We are looking for motivated students with some prior knowledge about MPC, optimization and programming in python. Knowledge of auto differentiation software (e.g. jax, pytorch, casadi) is a big plus. Please send your resume/CV (including lists of relevant projects/courses) and transcript of records in PDF format via email to rzuliani@ethz.ch and giacomo.delcaro@inspire.ch.

The project will investigate how day-ahead simulation results can be translated into concrete demand-shifting recommendations for residents, and whether modest incentives induce a measurable behavioural response. Based on this framing, various research directions can be explored (e.g. developing a prototype that informs residents). Potential research questions are: ▪ Can day-ahead push notifications and incentives measurably shift the demand profile of residents? ▪ Which incentive structures are most effective at sustaining participation over time without relying on sensitive data? ▪ Which KPIs best capture current performances and future projections and how do they support activities of residents? You would be hosted and supported by the Digital Twin & Informatics group at UESL with direct access to the NEST living-lab demonstrator, its residents, the ehub research platform and existing simulation models. Smart watches and supporting infrastructure for a field study are available, alongside a collaborative, supporting, international research environment.

You will explore how day-ahead simulation results can be translated into demand-shifting recommendations that residents act on, drawing on building simulation, data analysis, and behavioural research to design your approach and evaluate whether incentives drive measurable change.

Christian Schmid christian.schmid@empa.ch

With the increasing penetration of renewable energy resources, the system strength and voltage stiffness of power grids are declining, which challenges the stable operation of grid-connected converters. Grid-forming (GFM) technology is regarded as a promising solution for providing voltage and frequency support in weak grids. Besides stability, grid codes are increasingly specifying GFM capability requirements under large grid disturbances, such as inertial power response, active power response to phase jumps, and reactive power response to voltage dips. In addition, the dynamic performance of GFM converters, particularly response speed, is critical because it determines how quickly these services can be delivered. Faster responses can reduce the duration of deep voltage dips, accelerate post-fault voltage recovery, and improve fault-current contribution for protection. Therefore, the dynamic performance certification of GFM converters under large grid disturbances has become an important topic in ongoing grid-code discussions. The mainstream approach for estimating dynamic performance under large disturbances is based on time-domain simulation waveforms, which usually leads to case-by-case studies. Moreover, previous analytical studies have often relied on the voltage-source-behind-impedance assumption. Although this assumption provides useful insights, it neglects the control dynamics that shape the converter source voltage under grid disturbances. Therefore, more detailed analytical studies are needed to assess the feasibility of achieving the desired dynamic response performance under different control parameters and grid conditions.

This project addresses these gaps by analyzing the dynamic response of GFM converters under large grid disturbances, focusing on key parameters that govern the response, including control parameters and grid conditions. The study will also explore the tradeoff between stability and dynamic response based on the derived analytical dynamic performance indicators. The main tasks are as follows: - Introduce the detailed control structure of GFM converters and explain the functionality of control loops. - Analyze the influence of each control loop on the dynamic performance of GFM converters under grid disturbances. - Develop an analytical dynamic performance certification considering key control parameters and grid conditions. The tradeoff between stability and dynamic performance will be investigated. The derived certification will be validated in MATLAB/Simulink. - Apply the derived dynamic performance certification method to downstream tasks, such as GFM control design to mitigate the tradeoff between dynamic response and stability. The proposed improvement will be implemented and evaluated in MATLAB/Simulink.

Zheran Zeng, zengzh@ethz.ch Dr. Verena Häberle, verenhae@ethz.ch Competitive candidates should have a background in systems and control, power systems, and/or power electronics. Proficiency in MATLAB/Simulink is required for implementing and testing the proposed methods. Depending on the student’s interest, the project can be extended to hardware experimental validation. In addition, if the student is interested, support can be provided to publish the results in a conference paper.

This research will be carried out in parallel with a researcher at the Robotic Fabrication Lab (RFL) at the D-ARCH department, where the 3D printing of prototypes can be conducted. The student will build on an existing parametric workflow (Rhino + Grasshopper) to enable the integration of thermal and structural simulations with ANSYS software. Selected iterations may be 3D-printed at 1:1 scale, with performance validated through physical experiments including compression tests (structural) and a hotbox setup (thermal).

The main aim of this project is to develop a workflow for simulating the thermal and structural properties of 3D-printed ceramic components. Results of a parametric workflow should feed into the simulation software to establish relationships between set parameters and the performance of the components. Expected outcomes include: • An established workflow for enabling parametric design and simulation • Series of design iterations with thermal and structural simulation results • Possible experimental results with physical prototypes Successful results will be published as part of a scientific journal.

Please send your CV to lim@arch.ethz.ch. This project can be undertaken as an open project, a semester project or a master's thesis, and the scope will be adjusted accordingly. Feel free to send an email if you have any questions. More details will be given about the project during the initial discussions.

The goal of this project is to develop a robust startup procedure for multi-axis additive manufacturing, with a potential application on a robotic arm setup. The project focuses on geometric calibration. This includes mesh-based bed leveling and calibration of the rotational B and C axes. These steps aim to compensate for geometric errors such as non-planar build surfaces and kinematic misalignments. Thus, the project is expected to investigate a subset of the following problems: - Development of an automated mesh bed leveling procedure and compensation strategy; - Calibration of B and C axes and identification of kinematic errors in a 5-axis setup; - Integration of the calibration routines into an automated printer startup workflow; - Experimental evaluation of geometric accuracy, repeatability, and print quality under disturbances such as uneven bed surfaces or varying layer height. The work combines elements from control theory, robotics, and experimental system design, with a strong focus on real-world implementation.

The main goals of the project are as follows: - Design and implement an automated startup calibration pipeline, including bed leveling and axis calibration; - Improve geometric accuracy and repeatability of the 5-axis printing system; - Develop compensation strategies for measured geometric errors; - Evaluate the impact of automated calibration on print quality and robustness. - Deploy the developed procedure on a robotic arm printing setup.

Please email your resume/CV (including relevant projects) and transcript of records in PDF format to: iseckin@ethz.ch

This project is set to limited visibility by its publisher. To see the project description you need to log in at SiROP. Please follow these instructions: - Click link "Open this project..." below. - Log in to SiROP using your university login or create an account to see the details. If your affiliation is not created automatically, please follow these instructions: http://bit.ly/sirop-affiliate

This project is set to limited visibility by its publisher. To see the project description you need to log in at SiROP. Please follow these instructions: - Click link "Open this project..." below. - Log in to SiROP using your university login or create an account to see the details. If your affiliation is not created automatically, please follow these instructions: http://bit.ly/sirop-affiliate