Building things that perceive, act and learn.
Mostly harmless robotics engineer & computer scientist with a passion for mathematics.
Help me change my frame of reference
About
Hey there, I'm Magnus.
I am working to bring together robotics and modern AI, combining perception, control, and planning with the likes of Generative Models and Reinforcement Learning.
While Transformers, Diffusion Models and co. have revolutionized how we represent and interact with information through text, images, and videos, I believe that making robots truly intelligent requires a deeper grounding in first principles.
This grounding, I think, must come through inductive biases — mathematical structures that capture the patterns and regularities of our world, allowing learning systems to exploit them more effectively.
For me, the most promising frameworks to achieve this lie in the realms of Information Theory, Optimization and Lie Theory,
the languages of uncertainty, efficiency, and structure and symmetry.
Additionally, I believe that established methods such as Model Predictive Control, trajectory optimization, and massively parallel simulation remain essential foundations
— the challenge lies in integrating them with learning-based models to achieve reliable and generalizable robotic behavior.
Outside of research, where I am slowly but surely gradient descending into the local minimum many refer to as madness, I’m also an avid reader, traveler, and gym rat.
To learn more, pick a vector (category) to see which eigenvectors contribute (favorites):
News
Coming Soon ...
Contact
Email: mag.dierking@gmail.com
GitHub: github.com/magnusdierking
Paper
Ark
Open-source, Python-first robotics framework that provides a Gym-style interface for collecting data, training policies, and switching seamlessly between simulation and real-robot deployment. It includes reusable modules for control, SLAM, motion planning and visualization, and integrates natively with ROS to accelerate end-to-end robotics research.
View Paper
OpenPyro-A1
Open-source, low-cost bimanual half-humanoid robot designed for advanced manipulation research. It features a modular, repairable hardware design and supports coordinated two-handed tasks such as folding, cutting, and assembling. The platform enables teleoperation via a Meta Quest 3 and provides interfaces for learning-based controllers to support scalable, real-world robotics experimentation.
View Paper
More
Yeah, yeah, I'm working on it ...
University Projects
Model Tensor Planning
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
(Coming Soon)
Evasive Pick and Place
Developement of a robotic control system for a Franka Emika Panda in a dynamic simulation environment. Focused on enabling the robot to grasp and transport objects while evading randomly moving obstacles. The pipeline integrates real-time object tracking, advanced inverse kinematics with secondary objectives, and sampling-based motion planning to achieve adaptive, collision-free manipulation. Part of the "Intelligent Robotic Manipulation" course at the Technical University of Darmstadt.
View PDF
SO(3) Diffusion Model
Project report exploring the mathematical structures behind 3D rotations and rigid body transformations through Lie Theory. A deep dive into the underlying theory and its applications, presented in an intuitive and visual way. The project includes the implementation of a Diffusion Model on the manifold SO(3) and explores the impact of using non-Euclidean Brownian noise and geometric loss metrics. Part of a self-organized personal student project.
View PDF
Optimal Transport Seminar
Personal project report exploring the connection between information theory and optimal transport through the InfoOT framework. Studied how mutual information and entropic regularization can be combined to align data distributions, supported by implementations using Sinkhorn optimization and kernel density estimation. Part of a self-organized seminar following the book "Computational Optimal Transport" by Gabriel Peyré and Marco Cuturi.
View PDF
Medical Imaging VAE
Developement of a conditional VAE trained on a combination of medical imaging datasets. Focused on learning structured latent representations conditioned on metadata such as imaging modality or pathology. Explored data harmonization, cross-domain generalization, and visual synthesis to evaluate how conditioning improves generative performance across heterogeneous medical data. Part of the "Generative Models" course at the Technical University of Darmstadt.
Parallel Solution of Linear Systems arising in Domain Decomposition Methods
Implementation of a parallel solver for large-scale surface PDEs within an Isogeometric Analysis (IGA) framework. Developed a C++ implementation of the dual-primal IETI-DP algorithm based on Schur decomposition, enabling efficient parallel solutions of domain-decomposed problems. Verified the implementation on the Laplace–Beltrami operator using spherical harmonics, achieving the expected convergence behavior and strong runtime scaling.
View PDFMy Notes
After my Bachelor’s thesis, I began building a personal knowledge base using Obsidian — a habit that has grown with me through my Master’s, internship, and side projects.
Over time, it has evolved into a repository of thoughts, ideas, and notes spanning computer science, robotics, mathematics, and everything in between.
These notes are, and will always be, a work in progress — small attempts at understanding a universe that clearly didn’t come with documentation.
They’re less a collection of answers than a record of me trying to figure things out. Because of that, their quality ranges from near-empty stubs and quick sketches through half-formed ideas to blog-ready articles.
I also use several custom callouts and plugins that aren’t preserved when exporting, so some interface elements may look a little off.
When opening the vault in your browser, I recommend using light mode since many images have a white background. If your browser is set to dark mode, you might need to toggle between modes manually until you see callouts appearing under the first matrix definition.