Defense Day

`

1.1. Time and Place

The defense session in August will take place on August 30, 2024, 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/61786165009?pwd=Zkg0R2VvbWJHd28vbmJCa2RHdkZEQT09

Passcode: i6defense

1.2. Schedule

10:00 - 10:25 Qun Guo (MA Thesis)

Title: Saliency Object Detection Based Collision Avoidance for Mobile Robots in Unknown MultiObstacle Environments

Advisor: Yu Zhang

Keywords: 

Abstract:Obstacle avoidance is a critical capability for mobile robots operating in complex and dynamic environments. Traditional obstacle avoidance methods often struggle with the challenges posed by multiple dynamic obstacles, such as high computational complexity and difficulty in real-time avoidance of dangerous objects. To address these issues, this thesis proposes a novel obstacle avoidance method based on the E-FastSOD (Enhanced Fast Salient Object Detection) algorithm, which integrates salient object detection with depth camera perception. The proposed method leverages visual saliency information, depth data, and the E-MBC (Enhanced Minimal Bounding Circle) obstacle modeling approach to accurately identify and localize dangerous obstacles in the scene. The E-FastSOD algorithm incorporates depth and orientation saliency features to enhance the performance of dangerous obstacle detection. Salient regions are reconstructed in 3D space using depth information, and obstacle point clouds are processed using filtering techniques and an improved minimal bounding circle method to construct compact and regular obstacle models. The method also distinguishes between static and dynamic obstacles and predicts the motion of dynamic obstacles using a Kalman filter. The obtained obstacle information is integrated into the Type-II D-CBFs (Dynamic Control Barrier Functions) obstacle avoidance control strategy. Buffer D-CBFs are introduced to reduce computational complexity and improve real-time performance, and a safety controller is designed to ensure the robot’s safety during task execution. The effectiveness and superiority of the proposed obstacle avoidance method are comprehensively evaluated through simulations and real-world experiments, demonstrating its ability to enable safe and efficient navigation of mobile robots in complex dynamic environments. The main contributions of this thesis include: 1) A novel obstacle avoidance method based on the E-FastSOD algorithm, which integrates salient object detection with depth camera perception; 2) An Enhanced FastSOD algorithm and an improved minimal bounding circle method for efficient and accurate obstacle detection and modeling; 3) Integration of obstacle information into the Type-II D-CBFs obstacle avoidance control strategy and introduction of buffer zones to reduce computational complexity and improve real-time performance; and 4) Comprehensive evaluation of the proposed method through simulations and real-world experiments. This thesis provides new ideas and solutions for autonomous navigation of mobile robots in complex environments and lays the foundation for further research and development in the field of robot obstacle avoidance

10:25 - 10:50 Francesca D’Amico (MA Thesis)

Title: Trajectory Planning for Mobile Robotic Manipulation with Reinforcement Learning

Advisor: Josip Josifovski, Dr.-Ing Arne-Christoph Hildebrandt, Marco Baldini

Keywords: Reinforcement Learning, Trajectory Planning, Mobile Manipulation, Neural Networks

Abstract: The conventional approach to trajectory planning in industrial robotics relies heavily on predefined paths and necessitates human intervention for the manual selection of poses, which introduces significant limitations. This thesis addresses these limitations by integrating reinforcement learning (RL) into the trajectory planning process, thereby enabling the automatic selection of points and the generation of trajectories based on real-time observations of the surrounding environment. The proposed system is composed of two principal components. The initial component is the reinforcement learning (RL) algorithm, which determines the manipulator’s displacements by processing real-time environmental data. The second component is the motion generator, which is responsible for creating feasible paths for both the mobile base and the manipulator arm. The experimental results indicate that the RL models demonstrate strong performance in reaching designated goal positions and completing the mobile manipulation task. While there is room for improvement to fully match the robustness of baseline methods, these findings underscore the potential of RL to significantly enhance the adaptability and robustness of robotic trajectory planning.

10:50 - 11:15 Parikshit Bandhu Khana (MA Thesis)

Title: Real-time Panoptic Segmentation and Tracking of Anomalies in Wetlaid Non-Woven Sheets Production

Advisor: Kejia Chen

Keywords:

Abstract: Nonwoven microfiber sheets often suffer from defects that significantly impact their quality and the fiber structure. Conventional image processing techniques and adapted segmentationmodels struggle with the intricate structure, texture, and overlapping defects characteristic of these microfiber materials. This study explores panoptic segmentation models, which combine semantic and instance segmentation, to comprehensively address these challenges. Our approach aims to provide a universal solution for nonwoven defect segmentation, accurately identifying and categorizing various defect types within the fabric. By utilizing recent advancements in modeling overlapping object segmentation, we seek to improve the accuracy and robustness of defect detection at both microscopic and macroscopic levels. Furthermore, the superior segmentation capabilities of panoptic segmentation models allow for examining homogeneous and inhomogeneous regions, offering valuable insights for further research. This research highlights the potential of panoptic segmentation in advancing defect detection technologies for the nonwoven fabric industry and sets the stage for future exploration in this field.

11:15 - 11:40 Qinyang Xu (MA Thesis)

Title: Whisker-Inspired Tactile Sensing for Object Contour Extraction on Robot Manipulator

Advisor: Yixuan Dang

Keywords:  

Abstract: In nature, the mouse uses its whisker as an essential supplement to its senses, especially when optical information is unavailable. Inspired by this, this thesis explores one kind of design and usage of an artificial whisker sensor. This sensor is based on the Hall effect sensor, which is lightweight, small-scaled, and easily upgradeable. The sensor’s working parameters in the simulation and reality are calculated with calibration. This thesis then proposes a sensing algorithm to localize tip-contact points through readings from the deflections of the whiskers. In this way, it can return the location information of the obstacles it has contacted. The test results in simulation and reality show that the whisker can reflect contact points with sub-millimeter accuracy. Moreover, this sensor is accompanied by a controlling algorithm to continuously track the object in the desktop environment. This method includes a PI controller to adjust the running speed of the whisker and B-Spline function to calculate the slope of the contact surface and indicate the next gesture of the whisker as well. In this way, the whole contour of the obstacles on the desktop is sensed and reconstructed. The results show the sensor’s ability to accurately reproduce objects’ contours.

11:40 - 12:05 Rui Zhang (MA Thesis)

Title: Business Model Development for Autonomous Driving Industry in China: Economic Analysis, User Behavior Study and Policy Recommendations from Concept to Market Placement

Advisor: Liguo Zhou

Keywords: 

Abstract: This study explores the development of business models for autonomous driving technology and industry in the Chinese market, focusing on strategies that transition concepts into market-ready solutions. As the industry advances from L2 to L3 and moves toward L4, innovative models are crucial for commercial success. Through literature review and quantitative and qualitative analyses, four key strategies were proposed: (1) enhancing public awareness through scientific education; (2) diversifying industrial applications; (3) fostering government-enterprise-research cooperation through innovation platforms; and (4) updating relevant laws and regulations according to the development of the autonomous driving industry and China's specific conditions. These strategies were validated through collaboration with a startup, leading to significant breakthroughs in the Chinese market. Although these strategies are designed for China, they provide a framework that could inform the commercialization of autonomous driving and other emerging technologies globally. In summary, this study offers a comprehensive business model that integrates education, industrial strategy, collaboration, and regulation to support the growth of the autonomous driving industry.

12:05 - 12:30 Danya Liu (MA Thesis)

Title: Frequency Channel Attention Pattern Based Social-Transmotion: Advanced Trajectory Prediction

Advisor: Liguo Zhou  

Keywords: 

Abstract: The proposed network, FcaTransMotionNet, introduces a novel approach to human trajectory prediction by integrating a multi-spectral attention mechanism with a transformer-based architecture. This network effectively encodes spatio-temporal interactions through learned trajectory and identity embeddings, processed via a double ID encoder. A key innovation is the Multi-Spectral Attention Layer, which captures essential frequency components to enhance the model's ability to focus on relevant spatio-temporal patterns, ensuring robust trajectory predictions. The architecture is further strengthened by auxiliary transformer encoders that process both local and global information, contributing to the network's flexibility and adaptability, making it well-suited for complex, real-world scenarios in human motion prediction. To support this advanced architecture, a data loader framework was developed to handle complex, multi-person trajectory and pose datasets, specifically tailored for human motion prediction applications. This framework efficiently processes 3D joint coordinates and associated masks, ensuring data integrity through NaN value replacement, and supports flexible batching with variable sequence lengths via padding and masking techniques. It also includes specific implementations for datasets such as JTA, incorporating various visual cues to enhance model training. This robust and adaptable data loader is a critical component in the pipeline, enabling efficient management of large-scale, multi-modal datasets essential for accurate trajectory prediction. Experimental results validate the effectiveness of the proposed FcaTransMotionNet, demonstrating that the multi-spectral attention mechanism significantly outperforms the original social-transmotion model on the JTA dataset. This confirms that the new framework substantially improves trajectory prediction accuracy, marking a significant advancement over previous methods. 

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.