Student Projects
Safe Real-Time Online Learning for Autonomous Racing
Safely adapting to uncertain environments is a key requirement for robust autonomy. Robust Model Predictive Control (RMPC) optimizes decisions while explicitly accounting for uncertainty in the predictions, rendering it particularly well suited for safety-critical systems. In this thesis, you will develop and deploy a safe and efficient robust model predictive controller that can adapt its model under strict safety constraints.
Keywords
Autonomous Racing, Online Learning, Gaussian Process MPC, Robust Control
Labels
Master Thesis , ETH Zurich (ETHZ)
PLEASE LOG IN TO SEE DESCRIPTION
More information
Published since: 2026-07-10 , Earliest start: 2026-08-03
Applications limited to ETH Zurich
Organization Research Zeilinger
Hosts Lahr Amon
Topics Mathematical Sciences , Information, Computing and Communication Sciences , Engineering and Technology
Teaching Assistant: Few-Shot Adaptation of RL Policies on Real-World Impact Wrenches
This project investigates reinforcement learning for impact-wrench control. Reinforcement Learning (RL) is attractive here because it can learn a low-latency policy directly from interaction, but the impact dynamics, together with the relative scarcity of real-world data, make the problem hard. The project focuses on methods that target high performance from limited samples, such as offline RL, residual RL, and finetuning a simulation- or meta-trained prior, with the aim of quantifying on the physical tool what performance is reachable and how much additional data is needed to close the gap under real-world model mismatch.
Keywords
Reinforcement Learning, Applied RL
Labels
Machine Learning (PBL) , Student Assistant / HiWi , Robotics (PBL)
Description
Goal
Contact Details
More information
Published since: 2026-07-02 , Earliest start: 2026-07-01
Applications limited to Department of Mathematics , Department of Computer Science , Department of Information Technology and Electrical Engineering , Department of Mechanical and Process Engineering
Organization Center for Project-Based Learning D-ITET
Hosts Carron Andrea
Topics Information, Computing and Communication Sciences
Few-Shot Adaptation of RL Policies on Real-World Impact Wrenches
This project investigates reinforcement learning for impact-wrench control. Reinforcement Learning (RL) is attractive here because it can learn a low-latency policy directly from interaction, but the impact dynamics, together with the relative scarcity of real-world data, make the problem hard. The project focuses on methods that target high performance from limited samples, such as offline RL, residual RL, and finetuning a simulation- or meta-trained prior, with the aim of quantifying on the physical tool what performance is reachable and how much additional data is needed to close the gap under real-world model mismatch.
Keywords
Reinforcement Learning, Applied RL
Labels
Master Thesis , Machine Learning (PBL) , Robotics (PBL)
Description
Goal
Contact Details
More information
Published since: 2026-07-01 , Earliest start: 2026-07-01
Applications limited to Department of Mathematics , Department of Computer Science , Department of Information Technology and Electrical Engineering , Department of Mechanical and Process Engineering
Organization Center for Project-Based Learning D-ITET
Hosts Ghignone Edoardo
Topics Information, Computing and Communication Sciences
Structured Learning for MoE and Looped Transformers
This project investigates how sparse and low-rank structured learning can be extended to Mixture-of-Experts (MoE) models and looped transformers. Building on SALAAD, the goal is to study whether these architectures can be decomposed into shared low-rank components and sparse expert- or iteration-specific residuals. The project will analyze MoE expert specialization, routing behavior, and looped-transformer refinement dynamics, and evaluate whether such structure can improve interpretability, parameter efficiency, memory footprint, or inference-time computation without sacrificing task performance.
Keywords
Mixture-of-Experts, Looped Transformers, Sparse and Low-Rank Learning, Efficient Transformers, Model Compression
Labels
Master Thesis , ETH Zurich (ETHZ)
Contact Details
More information
Published since: 2026-06-30 , Earliest start: 2026-07-01 , Latest end: 2026-12-31
Organization Research Zeilinger
Hosts Ma Hao
Topics Mathematical Sciences , Information, Computing and Communication Sciences
SALAAD Beyond LLMs: Structured Sparse and Low-Rank Training for Multimodal Foundation Models
SALAAD is a structured training paradigm that introduces sparse and low-rank constraints during optimization to promote inherently compressible model representations without sacrificing task performance. While prior work has primarily focused on autoregressive language models, its applicability to broader model families remains unclear. In this project, we investigate whether SALAAD can generalize as a unified structured training framework across architectures and modalities, with a particular focus on multimodal foundation models such as vision-language and vision-language-action systems. These models introduce architectural heterogeneity and complex cross-modal interactions, posing new challenges for structured optimization. By combining insights from optimization, deep learning systems, and multimodal modeling, we aim to empirically and systematically evaluate the effectiveness of SALAAD in inducing compressible representations in such settings. Our findings seek to advance the understanding of structured training and its role in enabling efficient and elastic deployment of modern AI systems.
Keywords
Large Language Models, Foundation Models, Multimodal AI, Vision-Language Models, Vision-Language-Action Models, Deep Learning, Machine Learning, Artificial Intelligence, AI Systems, Representation Learning
Labels
Semester Project , Master Thesis , Other specific labels , ETH Zurich (ETHZ)
Description
Goal
Contact Details
More information
Published since: 2026-04-07 , Earliest start: 2026-04-07 , Latest end: 2026-12-31
Organization Research Zeilinger
Hosts Ma Hao
Topics Information, Computing and Communication Sciences , Engineering and Technology