AaronDsouza

About

I work where mechanics, control and software meet: modelling how a system moves, designing the controller that moves it, and testing that it does what the model promised.

1. 2026

   Daimler Truck AG

   Student production employee (holiday job)

   Wörth am Rhein

   • Worked with battery systems and their integration.
2. Apr – Sep 2024

   Airbus Defence and Space

   Bachelor's thesis: software development of ATTS post-processing tools

   Immenstaad, Lake Constance

   • Developed Python and MATLAB post-processing modules for the Airbus Thermal Tool Suite to automate the analysis and visualisation of spacecraft simulation results.
   • Migrated the platform from Python 2.7 to Python 3, resolved compatibility issues and integrated thermal tools from several thermal engineering departments.
   • Analysed the thermal and mechanical response of spacecraft models with Airbus engineering tools.
   • Validated outputs with tests so that simulation reports stayed reliable.
3. Jul – Dec 2023

   Airbus Defence and Space

   Engineering intern, Thermal Tool Suite development

   Immenstaad, Lake Constance

   • Simulated spacecraft thermal behaviour in Systema and ESATAN to support thermal engineering assessments.
   • Developed pre- and post-processing utilities in Python, MATLAB and Fortran for repeatable simulation data processing and reporting.
   • Built data visualisation in Python (Pandas, NumPy, Matplotlib) and MATLAB GUI tools on Linux.

   “His performance exceeded our expectations and requirements.”Airbus internship reference, translated from German

4. Mar – Dec 2022

   Hightech-Electronic GmbH

   Working student, production and testing

   Geldersheim

   • Assembled and tested battery modules in series and parallel configurations for Fresenius dialysis systems.
   • Inspected incoming Panasonic lithium-ion cells to verify their voltage was within specification.
   • Followed defined assembly and test procedures and documented every result, as medical technology requires.

• TUM Robot Modeling, Identification & Control Laboratory, Summer Semester 2026, graded 1.0 Robot modelling and parameter identification Kinematic and dynamic models of planar serial robots, with inertial parameters identified from excitation trajectories and used on the real lab setup. MATLABSimulink

  Challenge

  Model-based controllers need accurate kinematics and dynamics, but a robot's masses and inertias are rarely known precisely.

  My contribution

  Modelled planar serial robots with modified Denavit–Hartenberg parameters, homogeneous transformations, Jacobians and Lagrangian dynamics. Identified the inertial parameters using Fourier excitation trajectories, a minimal regressor and least-squares estimation.

  Methodology

  1. Forward kinematics from modified DH parameters
  2. Jacobians and Lagrangian equations of motion
  3. Dynamics written as linear in the parameters, reduced to a minimal set
  4. Fourier excitation trajectories
  5. Least-squares estimation of the parameters

  Tools

  MATLAB, Simulink

  Results

  The identified model improved the controller on the real robot. The lab course was graded 1.0, the best possible grade.

• TUM Robot Modeling, Identification & Control Laboratory, Summer Semester 2026, graded 1.0 PID and computed-torque control Joint trajectory tracking in MATLAB/Simulink, from PI/PID to model-based computed-torque control with state feedback and observers. MATLABSimulink

  Challenge

  Robot dynamics are nonlinear and coupled: each joint's motion changes the torque the others need. Fixed-gain PID handles this only approximately, especially at higher speeds.

  My contribution

  Implemented PI/PID and computed-torque controllers for joint trajectory tracking, along with state feedback and Luenberger and disturbance observers.

  Methodology

  Computed-torque control uses the dynamic model to cancel the nonlinear terms, leaving linear error dynamics set by the gains: τ = M̂(q) (q̈_d + K_D ė + K_P e) + Ĉ(q, q̇) q̇ + ĝ(q) The arm at the top of this page runs this control law live, with a deliberately imperfect model so a small tracking error stays visible. You can compare it with PID and adaptive control in the Lab.

  Tools

  MATLAB, Simulink

  Results

  Controllers and observers were tested on the real lab setup, as part of the lab course graded 1.0.

• TUM Robot Modeling, Identification & Control Laboratory, Summer Semester 2026, graded 1.0 Adaptive control and Lyapunov stability Lyapunov-based adaptive control that keeps tracking stable when the robot's parameters are uncertain, plus impedance control and collision observation. MATLABSimulink

  Challenge

  Computed-torque control degrades when the model is wrong, and the parameters can change, for example with a payload. Physical interaction also needs controlled compliance and reliable detection of contact.

  My contribution

  Implemented Lyapunov-based adaptive and impedance controllers, and force and collision observation.

  Methodology

  Choose a Lyapunov function that combines tracking error and parameter error, then derive the parameter update law so that its time derivative is never positive. The tracking error then converges while the parameter estimates adapt online.

  Tools

  MATLAB, Simulink

  Results

  Part of the same lab course, graded 1.0. Together with the projects above, it covers the full workflow from model through identification to a validated controller.

• Airbus Defence and Space, internship 2023 and Bachelor's thesis 2024 Post-processing tools for spacecraft thermal analysis Python, MATLAB and Fortran tools in the Airbus Thermal Tool Suite (ATTS) that turn raw thermal simulation results into checked plots and reports. PythonMATLABFortranSystemaESATANLinux

  Challenge

  Spacecraft thermal simulations produce large amounts of raw result data. Engineers spent a lot of time extracting, plotting and checking results by hand before they could assess a design.

  My contribution

  • Internship, Jul – Dec 2023. Simulated spacecraft thermal behaviour in Systema and ESATAN, including updating models when the design changed, and wrote pre- and post-processing utilities in Python, MATLAB and Fortran on Linux.
  • Bachelor's thesis, Apr – Sep 2024. Built post-processing modules in Python and MATLAB that read simulation output, analyse it and generate plots and reports. Migrated the suite from Python 2.7 to Python 3, resolved compatibility issues and integrated tools from several thermal engineering departments.

  Methodology

  Worked side by side with the thermal engineers who used the tools daily, adapted features to their feedback, and ran validation tests on the outputs so reports stayed accurate after each change.

  Tools

  Python (Pandas, NumPy, Matplotlib), MATLAB, Fortran, Linux, Git, Systema, ESATAN

  Results

  Repeatable, faster evaluation of simulation results and less manual work for the team. The internship reference states that my performance exceeded Airbus's expectations and requirements.

  Thesis

  “Research and Development: Migrating and Enhancing for Improved Efficiency and Accuracy”, supervised at Airbus and THWS

• TUM, Physics-Informed Machine Learning, Summer Semester 2026 Physics-informed neural network for a damped oscillator A neural network that reconstructs the full motion of a damped oscillator from only eight measurements by learning the governing differential equation, not just the data. PythonPyTorch

  Challenge

  Measurements are often expensive, but the physics is known. With only eight data points, a plain neural network cannot recover the motion between and beyond them.

  My contribution

  Implemented the network and a combined loss from eight measurements and the differential equation residuals at 200 collocation points, using automatic differentiation to evaluate the oscillator equation inside the loss.

  Methodology

  Loss = data fit at 8 measurements + ODE residual at 200 collocation pointsMinimising both terms gives a smooth, physically consistent solution x(t).

  Tools

  Python, PyTorch

• Personal projects on GitHub, Physical AI Classical vs learned control LQR and cascaded PID against PPO reinforcement learning on a cart-pole and a quadrotor, tested on the ways a real robot differs from its model. PythonNumPyLQRPPO

  Challenge

  A controller designed on a model has to keep working on a robot that differs from it: heavier loads, weaker or slower motors, delay, noise and wind.

  Projects and results

  • Cart-pole: LQR vs reinforcement learningAll controllers balance the nominal model. On 500 unseen robots, PPO with domain randomization succeeded in 87 % of runs, LQR in 54 %. Under 60 ms delay, PPO chattered at full force until a penalty on jerky actions cut its RMS force from 8.0 N to 0.3 N, at the cost of robustness (67 %).
  • Quadrotor: PID vs reinforcement learningOn the nominal drone, the tuned PID hovered as accurately as the randomized policy (7.7 cm RMS in 1 N gusts) with 4–5× calmer rotor commands. On 500 unseen drones, the randomized policy completed 62 % of flights, PID 30 %.

  Methodology

  Nonlinear models simulated with RK4 at 200 Hz, controllers at 50 Hz. PPO written from scratch in NumPy, trained with and without domain randomization. Every controller faces the same scenarios with the same random seed, and the write-ups list what would make the comparison fairer.

  Tools

  Python (NumPy, SciPy, Matplotlib), reinforcement learning (PPO)

  Try it

  The trained cart-pole policies run live in the Lab, against LQR.

• Personal projects on GitHub Python and MATLAB engineering tools Three open-source simulation and verification projects: motor speed control, battery state-of-charge estimation and requirements-based testing. PythonMATLABNumPy

  Challenge

  Work through common control and estimation problems end to end, with code anyone can rerun and results that reproduce.

  Projects and results

  • DC motor speed controlOpen loop vs P, PI and PID under a load step. Anti-windup cut the overshoot from 37 % to 0.2 %.
  • Battery state-of-charge estimationExtended Kalman filter vs Coulomb counting on a 70-minute drive cycle. The filter stayed within 2 %, while Coulomb counting ended 34 % off.
  • Cruise control verification13 requirements traced to automated tests with fault injection. All 13 pass; the tests caught a 2.9 km/h overshoot near top speed, which was then fixed.

  Methodology

  Model the plant, design the controller or estimator, then verify against stated requirements with automated, repeatable tests.

  Tools

  Python (NumPy, Matplotlib), MATLAB

Also

• Real-time object detection and trackingPersonal project, 2023. Detection and tracking from a live webcam with Haar cascades, contour extraction and bounding boxes in Python, OpenCV and NumPy.
• Sustainable propulsion concept for a historic ferryTHWS engineering project, 2023. Technical and economic trade studies of propulsion architectures, leading to a sustainable system concept.
• Project portfolio as a PDFDownload the six-page portfolio with diagrams of the main projects.

1. Since Oct 2025

   Technical University of Munich

   M.Sc. Mechatronics, Robotics and Biomechanical Engineering

   Munich

   • Robot Modeling, Identification & Control Laboratory, graded 1.0
   • Robotics, Robot Dynamics, Advanced Control, Nonlinear Control
   • Physics-Informed Machine Learning, Multidisciplinary Design Optimization
2. Oct 2020 – Oct 2024

   THWS, Technical University of Applied Sciences Würzburg-Schweinfurt

   B.Eng. Mechatronics

   Schweinfurt

   • Control Systems, System Theory, Design and Simulation of Mechatronic Systems, Actuators, Embedded Systems and Fieldbuses, Robotics and Production
   • Bachelor's thesis at Airbus Defence and Space: “Research and Development: Migrating and Enhancing for Improved Efficiency and Accuracy”
3. 2025

   Certificates

   Online specialisations

   Coursera

   • Machine Learning Specialization, DeepLearning.AI and Stanford University
   • Deep Learning Specialization, DeepLearning.AI
   • Google IT Automation with Python, Google

• RoboticsModelling, identification and control of robot manipulators.
• Robot learningLearned and classical controllers that hold up when the real robot differs from the model.
• Control systemsNonlinear and adaptive control with stability guarantees.
• AutomationAutomated systems that are reliable enough for industrial use.
• Model-based engineeringModels as the single reference from design to validation.
• Systems engineeringRequirements and traceability across mechanics, electronics and software.
• SimulationVerifying behaviour in simulation before it reaches hardware.