Defense Day

y`

1.1. Time and Place

The defense session in August will take place on September 26, 2025, Friday at 10:00 AM. You can join the Zoom meeting room using the information below. 

Join Zoom Meeting

https://tum-conf.zoom-x.de/j/69450849073?pwd=g1xjryZipuDDZ7aTZHDFb08D8CFzif.1

Passcode: i6defense

Hosted by Kejia Chen (kejia.chen@tum.de)

1.2. Schedule

10:00 - 10:20  Eisert, Moritz Oliver (BA Thesis)

Title: Whisker-Inspired Tactile Sensing for Dynamic Contact Localization on Biomimetic Rodent Robot

Advisor: Yixuan Dang

Keywords: 

Abstract: Robots navigating confined and cluttered spaces rely on accurate environment perception for collision-free movement. Traditional methods, such as lidar and vision-based sensors, provide high accuracy, but they are limited when dealing with translucent, low-light environments, and are expensive in manufacturing and computational resources. In such cases, tactile-based whisker sensors, bio-inspired by rodents, provide a robust way to perceive various surface features and structures in a proximal sensing range. This work introduces an enhanced state-based algorithm to improve contact point localisation along a magnetic transduced whisker sensor. A common challenge in tangential contact estimation is the non-injective mapping of instantaneous sensor measurements. We address this challenge by incorporating time-series sensor measurements and proprioceptive motion data. In addition, we propose a spatially continuous calibration model for both straight and tapered whiskers, which accounts for slip and friction effects along arbitrary trajectories and compensates for bending history relative to the relaxation state. Further, we propose an algorithm based on the Bernoulli Beam Theory to compute an initial state for an unscented Kalman filter upon contact. The filter incorporates intrinsic motion data from the robotic system into time series sensor measurements based on the generated calibration map, to estimate a contact point in real-time. Experimental results demonstrate up to a twofold improvement in convergence speed and reduced dynamic Euclidean error of 1mm and standard deviation of 0.5mm for various movement and deflection patterns. These results enable rodent-like robots to navigate cluttered spaces more safely and efficiently.

10:20 - 10:45 Zhiang Guo (MA Thesis)

Title: Trajectory-Based Dynamic Object Integration and Novel View Synthesis Using 4D Gaussian Splatting for Realistic Dataset Generation

Advisor: Rui Song

Keywords: 

Abstract: Understanding and synthesizing dynamic 3D environments is a critical challenge for
applications in simulation and dataset generation for autonomous driving. A novel
framework is proposed that explicitly integrates multi-object tracking (MOT) signals into
the Gaussian Splatting pipeline to achieve spatial-temporal awareness. Leveraging
segmentation masks, 3D objects positions, and timestamps, the method decomposes a
scene into static background and dynamic objects. A dedicated Tracking-Guided
Temporal Attention module enhances the representation by weighting spatial features
according to trajectory coherence, temporal dynamics, and object motion, effectively
bridging the gap between 3D object detection, tracking, and 4D Gaussian Splatting. By
advancing from 3D scene understanding to 4D aware scene modeling, this thesis
establishes a new paradigm for trajectory-guided scene reconstruction and novel view
synthesis based dataset generation.

10:45 - 11:10 Zongwei Zhang (MA Thesis)

Title: Active Assistance for Pathological Gait using Inverse Reinforcement Learning

Advisor: Zhenshan Bing

Keywords:

Abstract: Although many studies on exoskeleton control have been conducted, several challenges remain. For one, some approaches focus solely on the exoskeleton itself and neglect the interaction and cooperation between humans and machines, resulting in the human having to adapt their motion to the exoskeleton. Moreover, controls are often generated using traditional techniques that don't leverage user data. Although these methods can yield satisfactory results, one of the disadvantages is that they are often limited to providing support passively through predefined parameters without individualization, thus leading to poor adaptability across scenarios and latency in the motion. These disadvantages can reduce the effectiveness of the wearable or even cause injury during gait correction in medical applications. This thesis presents a control strategy for the exoskeleton that optimizes the generation of healthy gait by correcting pathological gait based on the user’s dynamic properties. The proposed solution employs inverse reinforcement learning to eliminate the need for manual reward engineering. Experiments are conducted in different environments to evaluate the feasibility of the approach, followed by training a unified policy capable of handling various impairments, including zero-shot cases. The final results demonstrate the potential of inverse reinforcement learning based methods for effective and adaptive gait rehabilitation.

11:10 - 11:30 Nicolas Pfitzinger (BA Thesis)

Title:From Zero-Shot to Fine-Tuned: Evaluating Mistral 7B for NL-to-SQL Generation on WikiSQL

Advisor: Fengjunjie Pan

Keywords:

Abstract: Natural language to SQL (NL-to-SQL) translation offers a user-friendly interface to relational databases, but commercial large language models impose high costs, data-transfer risks, and
limited customization. This thesis evaluates Mistral-7B, a 7 billion-parameter open-source model, for NL-to-SQL tasks. A controlled ablation framework is applied to the WikiSQL
benchmark, advancing from zero-shot prompting (1.62% execution accuracy) to schema injection (28.51%), followed by few-shot prompting (39.8%), and concluding with LoRAbased
parameter-efficient fine-tuning. Fine-tuning on 5,000 examples achieves 85.52% execution accuracy at an estimated training cost of $0.45, while full-dataset fine-tuning reaches
88.82%. Performance gains plateau by 5,000 samples, with residual errors comprising a ~1% syntax-error floor. Contributions include a reproducible implementation pipeline, a
taxonomy of error patterns, and practical guidance for balancing resource efficiency with accuracy in enterprise NL-to-SQL deployment. These results demonstrate that compact opensource
models can deliver near-commercial NL-to-SQL performance on affordable hardware, making them viable for enterprise adoption without reliance on proprietary APIs.

11:30 - 11:50 Malobrodskiy, Dennis (BA Thesis)

Title: Graphical User Interface for LLM AgenticWorkflow Visualization

Advisor: Fengjunjie Pan

Keywords:

Abstract: With the emergence and rapid improvement of Large Language Models (LLMs), workflows that utilize LLM functionalities in some of their agents have gained in importance. Existing solutions for agentic workflow software focus rather on functionality and simply a correct output than on the design of the graphical user interface (GUI). But the behaviour of these
LLM workflows is often opaque and difficult to reason about at scale. So naturally the need for a visualization of LLM agentic workflows appears which this thesis tries to address. It
presents a web-based GUI for visualizing and managing LLM agentic workflows, grounded in established HCI guidance on clarity, direct manipulation, hierarchical abstraction, error
prevention, and feedback. The interface renders workflows as node-and-lane diagrams with details-on-demand, real-time status indicators, and (potentially graphical) artifact inspection.
The latter aims to ease the maintenance of the workflow by allowing the user to inspect intermediate results in a human readable way. The workflow used to demonstrate the capabilities
of the GUI is an automotive code development workflow presented by Pan, Fengjunjie et al. at the chair for robotics, artifi- cial intelligence and real-time systems at the Techincal University
of Munich. It is just exemplary to show the capabilities of the GUI, the software could be used for other workflows in the future. Finally, various experiments on the GUI demonstrate
its benefits and validate the approach.

11:50 - 12:15 Xiyuan Wang (MA Thesis)

Title: Reinforcement Learning and Model Predictive Path Integral (MPPI) Control for Robot Manipulation Tasks

Advisor: Erdi Sayar

Keywords: RL, MPPI, Diffusion

Abstract: This thesis presents a framework that integrates diffusion-based reinforcement learning with model predictive path integral (MPPI) control for robotic manipulation. The method leverages diffusion models to generate subgoals that guide long-horizon planning, while MPPI ensures dynamically feasible and collision-aware trajectory execution.The framework is validated in simulation on a 7-DOF Franka Emika Panda robot across reaching and pushing tasks, including scenarios with obstacles. A neural signed distance field model is used to approximate obstacle geometry and provide collision cost feedback to the MPPI controller. Results show that the approach enables smoother and shorter trajectories in the reach task and achieves successful object push task.

12:15 - 12:45 Coffee break

12:45 - 13:05 Elli Hähnel (BA Thesis)

Title: Reliability of Template-Based Text Generation in Large Language Models

Advisor: Nico Reeb

Keywords:

Abstract: Real-world applications of Large Language Models (LLMs) in professional domains face significant reliability challenges despite their impressive language understanding and generation abilities. Continued limitations include inconsistent output format and the risk of hallucinations, especially in safety-critical applications. This thesis introduces a small dataset of four templates, covering the medical, legal, and business domains. Each template is paired with six corresponding source documents. The dataset is used as the basis to evaluate three models of different scales (Command R 7B, Llama 3.3 70B Instruct, and Claude 3.7 Sonnet) using two automated faithfulness metrics and a custom fine-grained evaluation framework for template adherence, which enables a detailed analysis of syntax failure patterns. The results show significant differences in performance between models.

13:05 - 13:25 Sebastian Feiertag (BA Thesis)

Title: Improving language-guided robotic grasping with uncertainty learning

Advisor: Hu Cao

Keywords: Grasping, Uncertainty learning

Abstract: State-of-the-art models for target-oriented robotic grasping are often deterministic, limiting their robustness when trained on noisy datasets with simple regression losses. This work reconsiders grasp synthesis as a probabilistic problem. We propose a model that predicts a full Gaussian distribution for each grasp parameter, trained with the Continuous Ranked
Probability Score. This proper scoring rule leverages the model’s predicted uncertainty as a rich training signal, enabling more robust learning. In parallel, we investigate architectural improvements, including task-aligned backbones (RegionCLIP), feature fusion networks (PANet), and vision-centric linear attention mechanisms (MSLA). Our experiments demonstrate that the probabilistic training objective provides the most significant performance gain, reducing the top-one grasp error rate by over 40% compared to the baseline. While architectural modifications also provide benefits, these gains are not additive when combined with the probabilistic framework. We conclude that for this task, improving the fundamental
learning objective can be more impactful than architectural refinement alone.

13:25 - 13:50 Kahraman, Eser Murat (MA Thesis)

Title: Enhancing Semantic Segmentation via Multi-Modal Fusion

Advisor: Hu Cao

Keywords: Semantic Segmentation, Fusion

Abstract: This thesis introduces robust fusion of visual RGB and depth features, and text-based guidance. By integrating textual modality into a state-of-the-art multimodal segmentation framework.

13:50 - 14:15 Yuchen Song (MA Thesis)

Title: IS-NODE: Learning Deep Incrementally Stable Dynamical Systems by Neural Ordinary Differential Equations

Advisor: Yingbai Hu

Keywords: Imitation Learning, Neural ODEs, Contraction Theory, Incremental Stability.

Abstract: We propose IS-NODE, a framework combining Neural ODEs with contraction theory to learn imitation motion policies with guaranteed incremental stability — ensuring exponential convergence of neighboring trajectories. It incorporates Lie algebra and VAE to handle manifold-valued orientations, significantly improving accuracy and robustness in execution.

14:15 - 14:40 Qinghua Fan (MA Thesis)

Title: Preference-Aware Attention Hypernetworks for Pareto Set Learning in Multi-Objective Robot Control

Advisor: Yingbai Hu

Keywords: Pareto Optimal Solutions, Multi-Objective Re inforcement Learning,  Hypernet.

Abstract: In this paper, we draw on the method of hypernet in MORL and SENet. This method can directly learn a continuous representation of the entire Pareto optimal set.   SEHyper-MORL  is a simple and efficient new method. It can quickly generate  good trained policies for different tasks and preferences. 

The I6 defense day is held monthly, usually on the last Friday of each month. The standard formats of talks are:

More information on preparing presentations can be found in the Thesis Submission Guidelines.