Manuel Philipp Vogel

Mitigating Covariate Shift in End-to-End Driving Using Gaussian Splatting | TUM (2025)

Ongoing Master’s Thesis about developing a pipeline to synthesize novel views of real-world driving scenes using Gaussian Splatting to augment training data and reduce covariate shift in end-to-end autonomous driving models

3D Reconstruction of Objects from a Single Source Image by Synthesizing Novel Views | TUM (2024)

Guided Research on 3D object generation from an image by learning precomputed attention maps. Grade: 1.0

Zero-Shot Completion of Partial 3D Shapes Using Multiview Consistent Inpainting | TUM (2024)

Part of the course Advanced Deep Learning for Computer Vision. Grade: 1.7

Non-Sequential Reinforcement Learning for Hard Exploration Problems | TUM (2023)

Improving methods to automatically generate expert demonstrations for sparse reward RL environments. Part of the course Advanced Deep Learning for Robotics. Grade: 1.3

Simulation-Based Autonomous Driving in Crowded City | TUM (2023)

Master’s project about developing an end-to-end autonomous driving approach in simulation. Grade: 1.0

Autonomous Mobile RC Unimog | RRLAB (2020)

Bachelor’s project about developing software for an RC Unimog to drive autonomously over a test track. Our team won the final competition where both unimogs drive simultaneous over the track, while trying to adhere to traffic rules and avoid each other.

LIDAR-GEIL: LIDAR GPU Exploitation in Light Simulations | SIMULTECH 2023

• Main author of a paper introducing a novel LiDAR simulation method. Accepted at SIMULTECH 2023.
• Using Unreal Engine’s GPU particles to simulate LiDAR beams, our approach enabled the simulation of millions of points per second and even outperformed depth-image-based approaches.
• DOI: 10.5220/0012085900003546

Method for the Targeted Warning of a Driver of an Oncoming Motor Vehicle | Porsche AG

• Inventor (100%) of a system that monitors the driver of an oncoming vehicle to assess potential danger and issue targeted warnings.
• Patent Link

System for Improving Simulated Representations of Real‐World Environments | Porsche AG

• Co-inventor (50%) of a method designed to enhance the realism of digital twins.
• The method aims at improving object locations, textures, and orientations in simulations based on comparisons with real-world photos.
• Patent Link

Method for Determining a Position of an Object Through Reflections | Porsche AG

• Inventor (100%) of a technique to improve the poses of an objects using reflections in monocular images, without requiring accurate depth estimation.
• Patent Link