Defense Day

`

1.1. Time and Place

The defense session in January will take place on February 28, 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/61786165009?pwd=Zkg0R2VvbWJHd28vbmJCa2RHdkZEQT09

Passcode: i6defense

Hosted by Jiajie Zhang (jiajie.zhang@tum.de) and Mingyu Liu (mingyu.liu@tum.de)

1.2. Schedule

10:00 - 10:25 Milica Kalanj (MA Thesis)

Title: Enhancing Robotic Teleoperation: Machine Learning for Force Feedback Prediction in Grasping

Advisor: Theodoros Papadopoulos, Josip Josifovski

Keywords: haptics, supervised learning, grasp modeling, force feedback

Abstract: Teleoperation has transformed human-robot interaction by enabling precise object manipulation in remote environments. The development of anthropomorphic robotic hands and haptic feedback has improved the realism and dexterity of teleoperated systems. However, realistic haptic feedback remains a challenge, as current systems often rely on heuristic or simulation-based methods with limited accuracy and realism. This thesis addresses these limitations using machine learning to predict motor torques and contact forces for generating haptic feedback. A grasp dataset for a Shadow Hand was generated and validated, serving as the foundation for torque prediction. A teleoperation pipeline was created to enable the users to control the Shadow Hand via the SenseGlove Nova and experience two modes of feedback: one from predicted torques and the other derived from simulation. A user study, which was conducted, showed that predicted feedback offered a more realistic and effective sensation for both seen and unseen objects. This thesis contributes to bridging the gap between human sensory input and robotic manipulation, demonstrating the potential of AI-driven feedback systems to enhance immersion and ownership during teleoperation.

10:25 - 10:45 Kaştan, Selim Mert (BA Thesis)

Title: Intelligent Safety Validation for Autonomous Vehicles

Advisor: Hu Cao

Keywords: Autonomous driving, Safety

Abstract: Ensuring safety of autonomous vehicles before widescale deployment is essential. However, traditional safety validation methdos such as real life testing is not only inefficent, but also very costly economically. Therefore, scenario based safety validation methods are promising alternatives for safety validation. However, the main problem in scenario based testing is that scenarios are written by humans. ChatScene tries to overcome this obstacle by automating the scenario generation using reasoning capabilities from large language models. However, algorithms used in autonomous driving technologies lack formal verification guarantees. Therefore, risk estimation while driving is life critical for autonomus vehicles. We bring an RSS-based risk estimation methodoly into ChatScene to further enhance its safety validation capabilities. By adding risk estimation, this approach enables AVs to assess potential dangers and improve decision-making. The research demonstrates that RSS can be utilized within ChatScene to quantify collision risk, offering a structured method for evaluating AV safety.

10:45 - 11:05 Maria Alejandra Sagastume Giron (BA Thesis)

Title: Neuromorphic Path Integration and Correction with Adaptive Sensory Fusion for Autonomous Driving Systems

Advisor: Genghang Zhuang

Keywords: Neuromorphic Path Integration, Sensory Fusion, Autonomous Driving

Abstract: Autonomous navigation is an absolutely essential feature for robotic systems, especially in the context of self-driving vehicles. Conventional techniques of localization and mapping mostly depend on computationally probabilistic methods. In contrast, neuromorphic computing provides an energy-efficient alternative inspired by neural mechanisms of biological navigation. This thesis explores a neuromorphic approach to path integration and correction for autonomous driving systems, using Spiking Neural Networks (SNNs) for real-time pose estimation. The primary challenge in path integration is the accumulation of drifts due to errors in pose estimation. An adaptive sensory fusion framework was proposed to periodically reset the pose estimates by combining visual odometry and input, hence reducing this. The model builds upon Continuous Attractor Network (CAN) and implements biologically plausible grid cell dynamics to maintain an internal representation of position. By employing Nengo and NengoDL, the network parameters were optimized and enhanced for computational efficiency. The approach was validated through simulations, comparing the performance of neuromorphic path integration with traditional Simultaneous Localization and Mapping (SLAM)-based techniques. Key performance metrics, such as Absolute Trajectory Error (ATE) and Relative Pose Error (RPE), show how well the adaptive corrective mechanism reduces drift. Results indicate that incorporating visual resets significantly enhances localization accuracy while maintaining a low computational footprint. This research contributes to the field of neuromorphic robotics by demonstrating the potential of biologically inspired navigation models for real-world applications. Further developments in neuromorphic SLAM and sensor fusion technologies are made possible by the scalable solution the suggested framework offers for autonomous systems functioning in demanding situations.

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.