Institute for Dynamic Systems and ControlOpen OpportunitiesCyberRunner is an AI robot whose task is to learn how to play the popular and widely accessible labyrinth marble game. The labyrinth is a game of physical skill whose goal is to steer a marble from a given start point to the end point. In doing so, the player must prevent the ball from falling into any of the holes that are present on the labyrinth board. The movement of the ball can be indirectly controlled by two knobs which change the orientation of the board. While it is a relatively straightforward game, it requires fine motor skills and spatial reasoning abilities, and, from experience, humans require a great amount of practice to become proficient at the game.
Using recent advances in model-based reinforcement learning techniques, CyberRunner is able to outperform the previously fastest recorded time, achieved by an extremely skilled human player, by over 6%. Moreover, it does so with only 6 hours of practice. We envision expanding the capabilities of CyberRunner through further research. Students will contribute to advancing the field and establishing CyberRunner as a real-world robotic benchmark.
Suitable projects in different areas are available for talented and motivated students. The project topics span model-based control, reinforcement learning, computer vision, and hardware design.
- Mechanical Engineering
- Master Thesis, Semester Project
| In this project, we want to explore possible extensions of predictive control barrier functions to the multi-agent setting. Predictive control barrier functions [1] allow certifying safety of a system in terms of constraint satisfaction and provide stability guarantees with respect to the set of safe states in case of initial feasibility. This allows augmenting any human or learning-based controller with closed-loop guarantees through a so-called safety filter [2] which is agnostic to the primary control objective. As current formulations are restricted to single agents, the goal is to investigate how this formulation can be extended for multi-agent applications and how the interactions between the agents can be exploited in order to reduce computational overhead. - Engineering and Technology, Systems Theory and Control
- Master Thesis
| The development of Large Language Models (LLMs), like ChatGPT and GPT-4, has influenced
the field of Natural Language Processing and Artificial Intelligence with their exceptional proficiency in comprehending and generating language, alongside their notable generalization and reasoning abilities. Consequently, recent research efforts have focused on leveraging the capabilities of LLMs to improve recommender systems. Recommender systems significantly influence human behavior by shaping users’ preferences, decision-making processes, and overall engagement with digital content. This project develops on the interpretation of recommender systems (controller) in feedback interaction with the users (system), [3]. By following a similar approach to [2], we will investigate how a careful integration of a LLM with a Model Predictive Control (MPC) framework can enhance recommender systems by ensuring accurate and adaptable recommendations while considering user preferences and constraints.
Understanding the influence of recommender systems over users behaviour and managing it effectively will be enhanced through the MPC framework, which offers a structured and interpretable approach to recommendation optimization. - Automotive Engineering, Computer Communications Networks, Electrical Engineering, Mechanical and Industrial Engineering
- ETH Zurich (ETHZ), Master Thesis, Semester Project
| Abstract
Reinforcement learning (RL) has achieved remarkable performance in various domains such as gaming, protein folding, and foundation models. However, efficiently applying RL to real-world applications like go-kart racing and urban driving presents significant challenges due to high-dimensional environments and the lack of structured task decomposition. This thesis proposes addressing these challenges through prioritized rewards and hierarchical task decomposition. By incorporating prioritized experience replay and dynamic reward shaping, the learning process focuses on critical experiences, enhancing efficiency. Hierarchical RL will break down complex tasks into manageable sub-tasks for better strategic planning and execution. The goal is to develop robust and adaptable RL agents capable of high performance in both racing and urban driving scenarios. The student will select a specific application domain, define benchmarks, and potentially conduct real-world testing. The outcomes are expected to contribute significantly to robotics research, with potential publications in top conferences and journals. Pre-requisites include a strong interest in machine learning, RL, optimization, robotics, and proficiency in Python. Prior experience in autonomous vehicles or robotics is a plus. - Automotive Engineering, Intelligent Robotics, Knowledge Representation and Machine Learning
- ETH Zurich (ETHZ), Master Thesis, Semester Project
| Our aim is to create an autonomous racing system capable of swiftly learning optimal racing strategies and navigating tracks more effectively (faster) than traditional methods and human drivers using RL.
- Information, Computing and Communication Sciences
- Master Thesis, Semester Project
| Mobility is typically self-optimized for a particular region to accommodate internal travel needs. However, as soon as one considers multiple, interacting regions (e.g., urban areas interacting with agglomerations, and agglomerations interacting with rural areas), important coordination issues occur, including scheduling mismatches, fleet allocations, and congestion peaks. In short, a mobility system composed of self-optimized mobility systems seems to often operate suboptimally.
In this project, we will investigate the idea of strategic interactions of future mobility stakeholders across heterogeneous regions, such as urban areas, agglomerations, and rural areas, leveraging techniques from network design, optimization, game theory, and policy making. - Automotive Engineering, Information, Computing and Communication Sciences, Mathematical Sciences, Mechanical and Industrial Engineering, Transport Engineering
- Master Thesis, Semester Project
| In this project, we want to explore the application of predictive stability filters for automotive applications. Predictive stability filters allow augmenting human or learning-based controllers such that safety in terms of constraint satisfaction as well as stability of a desired setpoint can be guaranteed. Such algorithms present possible solutions for automotive applications such as, e.g., lane keeping. - Engineering and Technology, Systems Theory and Control
- Master Thesis
| In many autonomous navigation applications, the robot must interact with the environment to learn and complete tasks. Furthermore, these applications are safety-critical, and crashes cannot be afforded. This necessitates the safe learning of the unknown environment in order to achieve the task objective (e.g., detecting a leak or mapping an area). For example, consider an application of safe exploration in a warehouse with a wheeled robot to identify the source of a gas leak. - Mechanical Engineering
- Master Thesis
| A key barrier hindering the swift introduction of autonomous vehicles (AVs) in real-world contexts is the challenge in establishing clear safety benchmarks. Specifically, the issue of systematically assessing both performance and safety remains a significant stumbling block within the industry.
This challenge is mainly twofold: Firstly, how can we identify an ideal scenario set to evaluate the vehicle's performance within a targeted Operational Design Domain (ODD) and what criteria would be useful in amplifying or paring down this set?
Secondly, how do we determine a substantial stopping criteria for the evaluation campaign, and what level of confidence should be attached to the observed performances? - Applied Statistics, Automotive Engineering, Intelligent Robotics, Other
- Master Thesis, Semester Project
| This project focuses on developing autonomous robots for synchronized performances on water. Equipped with kinetic water fountains, RGB lighting, and ultrasonic mist generators, the robots are designed to execute planned choreographies. The system utilizes robotics control, wireless communication, and positioning technologies to coordinate movements, and payload activation, facilitating complex pattern generation and synchronization. The objective is to advance the application of distributed robotic systems in creating structured and cohesive visual displays on water. - Arts, Engineering and Technology, Information, Computing and Communication Sciences
- Bachelor Thesis, Master Thesis, Semester Project
|
|