1.Mechanical Design and Integration for Multi-Domain Unmanned Platforms [Robotics] [Mechanical Design]

[BA/SA/MA]: [Proposal](contact: zy.zhang@tum.de, qian.huang@tum.de)  

This project explores the fascinating engineering challenge of building cross-domain unmanned platforms capable of seamlessly traversing air, ground, and aquatic environments. We aim to overcome the conflicting physical requirements of these domains—such as aerodynamic efficiency for flight, structural robustness for terrain, and hydrodynamic sealing for water—without relying on bulky, ad-hoc modifications. Building upon our fully functional 1st- and 2nd-generation prototypes, your goal is to design a unified, lightweight airframe featuring innovative morphing mechanisms (such as vector thrust articulation) and robust waterproofing systems. We are actively pushing the physical boundaries of hybrid robotic platforms. If you have your own ideas regarding reconfigurable structures, advanced materials, or rapid prototyping, let's chat!

2.Onboard Electronics and Embedded System Design for Multi-Domain Unmanned Platforms [Robotics] [Embedded Systems]

[BA/SA/MA]: [Proposal](contact: zy.zhang@tum.de, qian.huang@tum.de)  

This project explores the critical hardware-software interface that empowers complex, multi-domain robotic platforms. We aim to solve the bottleneck of fragmented electronics by building a unified onboard infrastructure capable of managing aerial, terrestrial, and marine operations simultaneously. Utilizing our existing advanced prototypes, your goal is to design custom, miniaturized PCBs for power management and sensor integration, develop robust embedded firmware (e.g., STM32) for state transitions, and seamlessly integrate open-source flight controllers (PX4/ArduPilot) tuned for unique multi-domain dynamics. We are actively defining the future of unified, plug-and-play electronic architectures for hybrid autonomous systems. If you have your own ideas regarding custom avionics, power distribution, or low-level robotic control, let's chat!

3.Autonomous Algorithms for Air-Ground-Marine Integrated Unmanned Platforms [Robotics] [Algorithm]

[BA/SA/MA]: [Proposal](contact: zy.zhang@tum.de, qian.huang@tum.de)  

This project explores the frontiers of autonomous intelligence for hybrid robotic platforms operating seamlessly across air, ground, and aquatic environments. We aim to solve the profound algorithmic complexities of cross-domain navigation—from safety-critical obstacle avoidance on unpredictable terrains to handling severe hydrodynamic disturbances during underwater task execution. Utilizing our state-of-the-art integrated prototypes , your goal is to develop robust algorithms for seamless locomotion mode transitions (e.g., flight-to-water) , agile full-degree-of-freedom underwater maneuvering , and multi-modal perception. We are actively defining the algorithmic "brain" for next-generation multi-domain autonomous systems. If you have your own ideas regarding reinforcement learning-based control, or trajectory optimization, let's chat!

Mentor: Yu Zhang, Qian Huang

1.[High Priority] Swimming Like a Dog: Mastering the "Dog Paddle" for Amphibious Quadruped Robot [Locomotion]

[BA/SA/MA]: [Proposal](contact: qian.huang@tum.de)  

This project explores the domain of Amphibious Locomotion Control and Sim-to-Real Transfer. We aim to solve the dynamic "clumsiness" legged robots face upon entering the water, where complex hydrodynamic forces—like buoyancy and drag—disrupt terrestrial walking gaits. By combining physics-based simulation environments (like Isaac Sim) with advanced control algorithms (Deep RL or MPC), you will build a locomotion framework that mathematically optimizes the rhythmic "dog paddle." Your goal is to enable our amphibious quadruped to efficiently generate forward thrust and maintain 3D stability in fluid environments, successfully deploying these virtual skills onto the physical robot.

 

2.[High Priority] Landing Like a Cat: Aerial Self-Righting for Amphibious Robots via Active Spine Control [Mechanical Design] [Locomotion Control]

[BA/SA/MA]: [Proposal](contact: qian.huang@tum.de)  
This project explores the fascinating intersection of Bio-inspired Dynamics and Aerial Control. We aim to solve the challenge of safe landing for agile robots without adding heavy, single-purpose reaction wheels (dead weight). Inspired by the "cat righting reflex," you will use Active Spine Control to manipulate the robot's angular momentum mid-air. Using physics simulation (MuJoCo/Isaac Sim) and trajectory optimization, your goal is to teach the robot to vigorously twist and reorient its body in milliseconds, ensuring a perfect four-point landing from any drop angle.
 

Mentor: Qian Huang

Note: 

Master’s thesis opportunities are continuously available until December 2028. New thesis topics are released on a rolling basis and are often assigned through internal referrals from former students, so this page may not always reflect the latest openings. For the most up-to-date thesis topics and opportunities, please visit my personal homepage.

1. Cross-Embodiment Lifelong Learning

The demo and recent publication on Nature Machine Intelligence related to this topic can be seen at: https://www.nature.com/articles/s42256-025-00983-2

[SA/MA]: Cross-embodiment lifelong learning leveraging common knowledge representation and continual world model (contact: y.meng(at)tum.de)

In the process of lifelong learning for robots, cross-embodiment knowledge transfer and reuse are often overlooked. This leads to overfitting to specific dynamic embodiments and a lack of generalization. However, embodiments with similar dynamic characteristics hold the potential for direct task knowledge transfer and generalization. Therefore, this project focuses on cross-embodiment knowledge co-representation and transfer during lifelong learning. We propose to use a Mixture of Experts (MoE) architecture to align the inputs and outputs of different embodiments and to construct a shared representation in the latent space. At the same time, we will incorporate a world model to enable the latent space to predict states independent of specific dynamics, aiming to improve the cross-embodiment transfer performance in lifelong learning.

Keywords: Lifelong Reinforcement Learning, Cross-Embodiment, Shared Latent Representation, Transformer, Mixtrue of Experts, Generative Experience Replay, Diffusion Modeling

2. Large Language Models Empowered Humanoid Manipulation

[SA/MA]: Recursive Self-Improved Harness Framework for Humanoid Dual-Arm Manipulation (contact: y.meng(at)tum.de)

Large language models (LLMs) have demonstrated strong potential for autonomous robotic manipulation through code generation and structured tool use. However, reliable long-horizon humanoid manipulation remains challenging due to complex dual-arm coordination, dynamic environments, and execution failures. This thesis aims to develop a recursive self-improving harness framework for humanoid dual-arm manipulation, integrating LLM-based task planning, modular robotic APIs, and closed-loop execution feedback. The framework will enable LLM agents to generate and execute manipulation programs while monitoring execution outcomes, identifying failures, and autonomously refining action strategies. A key focus is on recursive self-improvement, allowing the system to learn from execution failures, iteratively refine generated code, and improve reusable manipulation skills. Students will contribute to the development of the agentic harness architecture, perception and manipulation APIs, and feedback-driven failure recovery mechanisms. The framework will be validated through real-world humanoid manipulation tasks involving dual-arm coordination and autonomous failure recovery. Strong Python programming skills are required, while experience with ROS, LLM agents, or robotic manipulation is preferred.

Keywords: LLM Agents, Code Generation, Agentic Harness, Recursive Self-Improvement, Humanoid Manipulation

[SA/MA]: Vision-Language-Action Model with Humanoid DexHand Manipulation (contact: y.meng(at)tum.de)

Recent progress in Vision–Language–Action (VLA) models and embodied intelligence has brought the robotics community to what many consider the “pre-GPT moment” for robots. However, current VLA models face significant limitations. On the data side, most existing VLA datasets are designed for simple single-arm robots with two-finger grippers, while large-scale datasets for humanoid robots equipped with dexterous hands remain scarce. Meanwhile, internet-scale human videos, an abundant source of prior knowledge for robotic learning, are still underexplored. Correspondingly, existing VLA architectures are not well adapted to highly articulated humanoid robots with dexterous hands, which often involve more than 50 degrees of freedom. This thesis project proposes to leverage human interaction videos for world-model learning, combined with small-scale robot data for hierarchical VLA fine-tuning, to enable dexterous manipulation on humanoid robots. This is an open research topic, and students with strong academic performance (GPA<2.0) or prior research/publication experience are encouraged to apply. Outstanding students may have opportunities to co-author publications.

Keywords: Vision-Language-Action Models, Humanoid Manipulation, Dex Hand, Imitation Learning.

3. RL Post-Training for Humanoid Manipulation

[SA/MA]: RL Post-Training for Failure-Aware Humanoid Manipulation (contact: y.meng(at)tum.de)

We are recruiting students to join a research project on reinforcement learning (RL) post-training for robust humanoid manipulation, with a particular focus on dual-arm decoupled policy optimization and failure recovery. This project aims to enhance the robustness and adaptability of robotic manipulation policies through RL-based post-training on real humanoid robots. Building upon pre-trained robotic policies, we will investigate decoupled learning strategies that enable independent optimization of dual-arm manipulation behaviors while preserving coordinated execution. In addition, we will develop failure-aware recovery mechanisms that allow robots to detect and correct manipulation failures through targeted policy refinement and online interaction. Students will work on the full pipeline, including VR-based teleoperation and data collection, VLA/VA model fine-tuning, RL post-training, and real-world deployment on humanoid platforms. We are looking for students with solid Python programming skills and experience in ROS, while prior experience in reinforcement learning (RL) or imitation learning (IL) is highly preferred. The position requires a minimum commitment of 6 months full-time and offers hands-on experience with state-of-the-art embodied AI systems, real-world robotic learning, and opportunities for academic publication.

Keywords: Humanoid Manipulation, RL Post-Training

Applicants interest in above topics are expected to have:

• Proficiency in Python, Linux OS and a solid understanding of deep learning.
• Good English communication and academic writing skills.
• Good academic performance (GPA ≤ 2.3).
• Availability for full-time work at least 6 months.
• Prior experience with reinforcement learning/ imitation learning is highly desirable and will be considered an advantage.

Additionally, we have several ongoing projects and we also accept open proposals or ideas with yourselves. Students who achieve outstanding results will have the opportunity to publish papers with us as co-authors in international top-tier journals/conferences. We welcome interested students to reach out at any time: Yuan Meng

1.Collaborative Underground Mapping for Cross-Scale Heterogeneous Robots

[BA/SA/MA]: [Proposal]

Underground environments pose significant challenges for autonomous mapping due to complex terrains, making single-robot coverage insufficient. To overcome this, cross-scale heterogeneous robot teams offer a promising solution by combining large robots for global localization and computing with small robots for exploring narrow, high-risk gaps. However, challenges like sensor fusion, communication limits, and payload constraints persist. This project proposes a mothership-child collaborative exploration system. The mothership acts as a centralized sensing and computing hub, guiding child robots that explore inaccessible areas. Ultimately, the project aims to develop a framework for efficient multi-robot cooperation and accurate map construction.

2.Locomotion of a Rat Robot with 3-DOF Legs in Curved and Narrow Spaces

[BA/SA/MA]: [Proposal]

Small-scale quadruped rat robots offer significant potential for exploring confined areas, but balancing a compact design with maneuverability remains a challenge. While traditional biomimetic spines require too much clearance for narrow spaces , a previously designed 3-DOF limb structure successfully optimized linear movement. However, how this structure handles complex turning is still untested. Because tight L-shaped or U-shaped tunnels impose strict turning radius limits , improving tight-turn passability and overcoming motion singularities—without sacrificing the robot's compact size—is essential for real-world applications.

3.Whole-Body Balance Control for Injured Quadrupedal Robots

[BA/SA/MA]: [Proposal]

Small-scale quadruped rat robots offer significant potential for exploring confined areas, but balancing a compact design with maneuverability remains a challenge. While traditional biomimetic spines require too much clearance for narrow spaces , a previously designed 3-DOF limb structure successfully optimized linear movement. However, how this structure handles complex turning is still untested. Because tight L-shaped or U-shaped tunnels impose strict turning radius limits , improving tight-turn passability and overcoming motion singularities—without sacrificing the robot's compact size—is essential for real-world applications.

4.Pathological Gait Generation for Injured Quadruped Robots Using Reinforcement Learning

[BA/SA/MA]: [Proposal]

While quadruped robots are highly adaptable, their complex architecture makes them susceptible to hardware failures like limb loss, which render standard control strategies ineffective. Preserving mobility requires generating a stable tripedal pathological gait, but coordinating the remaining limbs involves highly nonlinear constraints that are difficult for conventional model-based methods to handle. Reinforcement learning (RL) offers a solution by autonomously discovering compensatory coordination through simulation, providing a highly resilient approach for fault-tolerant control and gait generation under complex damage conditions.

5.Pathological gait generation for rat robot with spine-based damage control

[BA/SA/MA]: [Proposal]

Quadruped robots rely on precise foothold selection for mobility. However, irreparable limb failure disrupts this ability, causing asymmetric motion and severe issues with movement direction and dynamic balance. Consequently, damage control mechanisms are rarely applied to quadrupeds compared to hexapods. Inspired by animals that utilize their spines for movement direction and balance , this research explores damage control using a rat robot equipped with a flexible spine. The objective is to generate pathological gaits that enable the robot to maintain balance and continue task execution despite irreversible limb damage.

6.Design of a Wheel-Legged Rat Robot for Enhanced Mobility

[BA/SA/MA]: [Proposal]

While rat robots are ideal for exploring confined spaces, their small size inherently restricts their actuation, payload, and obstacle-clearing capabilities. Relying purely on legged locomotion often results in low speed, high energy consumption, and poor adaptability in mixed environments. To address this, converting rat robots into wheel-legged platforms offers a highly effective solution. This hybrid approach combines the speed and energy efficiency of wheeled motion on flat surfaces with the obstacle-negotiation advantages of legged systems, ultimately overcoming size-induced constraints and significantly enhancing the robot's robustness and practical utility.

Mentor: Yuhong Huang(yuhong.huang@tum.de)  Qian Huang(qian.huang@tum.de)