Safe reinforcement learning
PPO with Control Barrier Functions, curriculum learning, and domain randomization, taken from simulation to a real quadrotor.
Available for remote engagements · Worldwide
Hi, I'm Oguzhan —
I build robots that learn — and stay safe doing it.
Robot learning & control engineer and Physical AI researcher at TUM. I work on reinforcement learning, diffusion and flow-matching policies, and continual learning, and I bring 4+ years of industrial control experience to it. I take on remote engagements as an independent engineer, on a clean B2B basis with no payroll or employer overhead on your side.
How I work · Engagement Model
Senior engineering through a clean B2B engagement — no payroll, no employer contributions, no permanent fixed cost — on a structure Turkish law explicitly supports.
No employer overhead
You pay for the work — not the social charges, benefits, paid leave, or severance layered on a local salary.
Senior level, competitive rate
Türkiye's cost base and tax framework let me deliver senior-grade engineering well below a comparable local hire's total cost.
You pay for output
Scoped, deliverable-based B2B billing. You commit to a project, not a permanent headcount.
Legitimate & contract-backed
A registered business with its own tax identity. Scope, NDA, and IP transfer set out in a written services agreement.
01 — About
Where control theory meets robot learning
I work on robot learning: getting robots to pick up skills from data and experience while keeping the safety and reliability that real hardware needs. Today that means reinforcement learning, diffusion and flow-matching policies, and continual learning on humanoids, quadrupeds, robot arms, and drones.
I came to it from the control side. I spent 4+ years at AVL building model-based control software for electric vehicles, hybrid systems, and autonomous platforms to ASPICE and MISRA standards, so I start from the dynamics, put constraints on the learner, and only trust results once they hold up on the real robot.
At TUM I'm a Physical AI Research Assistant working with the Unitree G1 humanoid, B2W quadruped, and AgileX arms. My master's thesis is on continual learning: how manipulation policies can keep learning new tasks without forgetting old ones. Before that, I combined RL with Control Barrier Functions to recover a real quadrotor from fully inverted states.
I also learn new things quickly. I moved from automotive control software into state-of-the-art robot learning and use each project to add a new method or platform. Research moves fast, and I keep up with it.
I'm also an independent engineer available for remote engagements worldwide in robotics, control, and machine learning, engaged on a clean B2B basis rather than as an employee. See how I work →
- 4+Years industry experience
- 3Robot-learning research roles at TUM
- 1Patent · 1 paper submitted
- 6+Robotics projects
02 — Research
Robot learning research
My research covers how robots learn, adapt, and stay safe, with each idea tested on real hardware.
Generative & imitation policies
Diffusion and flow-matching policies for manipulation, plus imitation learning from motion-capture data for humanoid gait.
Continual learning
Learning sequential manipulation tasks without catastrophic forgetting: replay-free methods and compositional generalization.
Foundation models & physical AI
VLMs and LLMs for high-level reasoning on top of classical control, deployed on humanoids, quadrupeds, robot arms, and race cars.
Physical AI Research Assistant
Autonomous Vehicle Systems Lab, TUM
- Working hands-on with the Unitree G1 humanoid, B2W quadruped, and AgileX robot arms.
- Supporting SAP in leveraging large-scale business context data for robotics applications.
Master Thesis Student
Autonomous Vehicle Systems Lab, TUM
- Developing continual learning methods for multi-task robotic manipulation.
- Investigating catastrophic forgetting and interference in flow-matching and diffusion policies.
- Evaluating replay-free learning and compositional generalization across sequential manipulation tasks.
Semester Thesis Student
Autonomous Aerial Systems Lab, TUM Code Project page
- Built a safety-critical RL framework combining PPO with Control Barrier Functions for safe quadrotor recovery and stabilization.
- Derived nonlinear quadrotor dynamics, Lie-derivative-based CBF constraints, and real-time safety-filtering pipelines.
- Designed curriculum learning and reward shaping (dual-scale Cauchy rewards with goal-directed velocity objectives) to improve recovery from arbitrary initial conditions.
- Added domain randomization, parameter-uncertainty modeling, and calibrated disturbance injection for robustness and sim-to-real transfer.
- Extended Flightmare with Stable-Baselines3 integration and disturbance models; built an NMPC baseline for comparison.
- Deployed the learned policy on a real Agilicious quadrotor — robust recovery from fully inverted states; results contributed to a paper submitted to an international controls conference.
03 — Experience
Industry experience
Working Student
AVL Software and Functions
- Continued my model-based control development responsibilities part-time alongside my master's studies.
Control Development Engineer
AVL
- Developed model-based control software for electric vehicles, hybrid systems, autonomous robotics, and e-mobility platforms.
- Built application-layer software in MATLAB/Simulink compliant with ASPICE and MISRA standards.
- Designed and maintained safety-critical control functions and algorithms — protection logic, thermal management, contactor control, diagnostics, and vehicle communication.
- Executed Model-in-the-Loop (MiL) and Software-in-the-Loop (SiL) testing and supported HiL integration and validation.
- Managed component-level requirements and tests; helped define and improve modular, reusable software architecture.
- Built Python tooling for model calibration and prototyped deep-learning algorithms for state estimation.
- Collaborated with cross-functional, multi-disciplinary teams across AVL locations worldwide — a fully distributed, remote-first way of working.
Working Student
Siemens
- Supported after-sales operations and customer incident resolution.
- Troubleshot non-conformities in medium- and low-voltage switchgear (electrical & mechanical).
- Coordinated stakeholders and subject-matter experts across operations; managed procurement, dispatch, and customs of critical materials.
- Tracked and reported after-sales non-conformities to improve quality requirements.
04 — Projects
Selected projects
Safe RL for Quadrotor Recovery Paper
PPO policy with a Control Barrier Function safety filter that recovers a quadrotor from fully inverted states. Beats an NMPC baseline on extreme starts and was deployed on a real Agilicious drone.
Foundation-Model Autonomous Driving on F1TENTH
Full ROS 2 autonomous-driving stack on NVIDIA Jetson Nano integrating perception, localization, planning, and control. Combined VLMs and LLMs for high-level decisions with a Stanley controller for lane tracking, guarded by a finite-state-machine safety layer. Validated end-to-end on real hardware.
Humanoid Gait Imitation with RL
MuJoCo RL framework for humanoid locomotion using OpenSim inverse kinematics from motion-capture data for imitation learning, with reward shaping for stable, physically plausible gait.
Read moreAutonomous UAV Cave Exploration & Navigation
ROS 2 quadrotor for 3D cave exploration in Unity. OctoMap occupancy mapping, frontier-based exploration, sampling-based motion planning, and a geometric SE(3) controller for aggressive 3D trajectory tracking.
Mobile Robot for Intralogistics 1st place
Raspberry Pi LEGO robot for warehouse navigation. Dijkstra global planning, OpenCV perception, PID motor control for differential drive, and MQTT-based task allocation — won 1st place in the course challenge.
Read moreEmbedded Ball-and-Plate PD Control
Real-time PD control on a dsPIC33F for closed-loop circular trajectory tracking. Low-level embedded software using registers, interrupts, ADC sensing, and PWM dual-servo actuation — validated on real hardware.
Read morePatent — Charging Current Limit for Rechargeable Batteries
Contributed to a hybrid fast-charging framework for lithium-ion batteries that combines physics-based electrochemical modeling with data-driven prediction and control.
View on Google Patents
05 — Skills & Education
What I work with
Technical skills
- Python
- MATLAB & Simulink
- C++
- Object-Oriented Programming
- ROS / ROS 2
- Linux
- Git
- PTC Windchill
- Codebeamer
Learning & simulation
- PyTorch
- Stable-Baselines3
- MuJoCo
- Flightmare
- OpenCV
Domains
- Reinforcement Learning
- Imitation Learning
- Diffusion & Flow-Matching Policies
- Continual Learning
- Sim-to-Real
- VLM / LLM for Robotics
- Model-Based Control
- Control Barrier Functions
- NMPC
- Safety-Critical Systems
- Embedded Systems
- ASPICE / MISRA
Languages
- English — C1 (IELTS 7.0)
- Turkish — Native
Education
Technical University of Munich
Oct 2024 — Jan 2027M.Sc. Mechatronics and Robotics · Munich, Germany
Yıldız Technical University
2018 — 2022B.Sc. Electrical Engineering · Istanbul, Turkey
GPA 3.57/4.0 (German equivalent 1.6)
06 — Remote work
How I work remotely
I've spent years collaborating across globally distributed teams, and I'm set up to be effective from day one in a remote-first engagement — wherever your team is.
Distributed by default
At AVL I worked daily with cross-functional teams across multiple countries and sites — async collaboration and remote delivery are how I'm used to operating.
Flexible hours & overlap
Based in Munich (CET). Happy to flex my schedule to keep solid daily overlap with teams across Europe, North America, and beyond.
Clear async communication
I write things down — concise updates, documentation, and well-scoped tickets — so progress stays visible and decisions are easy to follow without constant meetings.
Self-directed delivery
Across industry and research I've taken projects from ambiguity to working hardware independently — comfortable owning outcomes with minimal supervision.
07 — Contact
Let's scope your next engagement.
I take on remote engineering engagements in robotics, control, and machine learning. Tell me about the project and I'll come back with a clear scope, deliverables, and a rate. The fastest way to reach me is email.