Publications
Journal Papers
Camera-Orientation Effects in Robotic Viewpoint Acquisition for Feed-Forward 3D Reconstruction in Manufacturing Inspection Workcells
Adam Michael Altenbuchner, Richard Matthias Hartisch, Taha Mohammed, Ziad Abouhalawa, Artem Balatsiuk, Aziz Louati, Haroun Lallouche, Jörg Krüger.
The International Journal of Advanced Manufacturing Technology (Springer) — under review, submitted June 2026.
Abstract. This paper evaluates camera orientation as a controllable acquisition variable in a simulation-based robotic inspection pipeline combining UR5e eye-in-hand viewpoint execution in NVIDIA Isaac Sim, feed-forward reconstruction (VGGT, Fast3R, SAM3D), point-cloud preprocessing, metric-scale correction, and registration-based comparison against reference geometry. VGGT achieved the highest mean registration fitness (0.93) and lowest mean inlier RMSE (0.002 m) across a 28-object evaluation set. Across five reproducible scan patterns, object-pointing camera orientation increased mean registration fitness from 0.68 to 0.79 and reduced mean inlier RMSE from 0.035 m to 0.022 m versus fixed-downward orientation. A coverage-oriented RL module (PPO, Isaac Lab) required dense proximity reward shaping before sparse coverage rewards supported reliable exploration (45.2% success vs 0.4%).
My contribution: Data preparation and validation — dataset curation, ground-truth geometry preparation, and evaluation pipeline support across the 28-object registration protocol.
Technical White Papers
Path Matters: Reinforcement-Learning-Driven Camera Trajectory Optimization for Robotic 3D Reconstruction
Taha Mohammed. TU Berlin, 2026.
Technical white paper documenting the MDP formulation, PPO training, reward shaping methodology, experimental results across multiple training seeds (with variance), and explicit limitations. Reports 45.2% task success versus 0.4% random-trajectory baseline under matched step budgets, with mean VGGT fitness 0.93 on held-out evaluation objects.