Control systems, mechatronics, autonomous driving. I build the rig, write the firmware that runs it, model the plant, and measure how far the model was off.
The repositories carry the measurements as well as the code: step responses, identified parameters, and the runs that went wrong along with what changed afterwards.
KR3 cell platform (private, in progress) — a simulation-first ROS 2 Humble platform for the university lab's KUKA KR 3 R540 on a KR C4 compact. One launch file, one argument: the same applications run against a mock arm (L0), Gazebo physics (L1), a simulator of KUKA's RSI protocol (L2) and the real controller (L3), which now runs in the loop. C++ only, with an RSI protocol stack in C and its own ros2_control driver.
- Applications — MoveIt 2 and Servo teleoperation, randomized pick-and-place, palletizing, conveyor picking, welding and dispensing paths, all gated by a cell supervisor (modes, watchdog, speed limit) that sits in front of every motion.
- Physics, not bookkeeping — at L1 every part in the planning scene is mirrored into Gazebo, carried while grasped, then left to fall and settle; the measured rest pose goes back into the planner. The first runs found what L0 could not: a brick-bond pallet pattern that is unstable at that pitch, a 3 mm release clearance the stack height had absorbed, and a carried body accumulating velocity under gravity.
- Every run recorded — MCAP bags and an SQLite experiment database tied by run id, with the git commit, model and controller-config hashes, per-joint tracking error and application metrics. Runs whose setup failed are flagged and kept out of the statistics; a daily report is generated from the database. 222 tests (CI and local runs).
- On the real arm — the KR C4 in the loop over RSI: kinematics within 0.009 mm on 20 poses, ~760 moves driven from the cell PC (in automatic mode behind the cell's guard), every one within 0.002°, ~830 000 real-time packets with none missed, fault stops measured on steel. The recordings identified a ~31 ms controller dead time; with it the digital twin matches the real arm within 0.007° RMS on sessions not used for the fit.
- Next — the same calibration for the physics level, wider motion ranges over RSI, gripper selection and integration.
Bosch Future Mobility Challenge 2026 — team OdyssiX. A 1:10 scale car driving an urban circuit unaided: lane keeping, signs and traffic lights, intersections, parking. Two repositories, talking to each other over micro-ROS at 460 800 baud.
| BFMC2026-Brain | Jetson Orin Nano, ROS 2 Humble, C++. Perception (lane geometry, free space, detection and tracking), planning (behaviour, global and local paths, mission supervisor, trajectory generation), EKF localisation and drift estimation, a safety layer with watchdog and emergency stop, and operator UI and telemetry. colcon, CycloneDDS, Docker for the driver containers. |
| BFMC2026-Embedded | STM32F401RE Nucleo, C, FreeRTOS. BNO085 IMU read at 100 Hz over I²C on an interrupt, published over micro-ROS; TIM2 PWM out to the steering servo and a Quicrun ESC against a bench-calibrated map; a 500 ms command-silence deadman that returns both to neutral. 1 kHz loop, minimal ISRs, lock-free UART ring. |
thin-structure-segmentation-evaluation — mean-IoU under-reports thin structures like lane lines, and the size of the error follows from the geometry. Derivation, a benchmark against clDice, 62 tests. An evaluation study, not a new model.
A rig, a model of it, and the gap between the two.
| steval-edukit | Rotary inverted pendulum on an STM32F401RE. Bare-metal C at 500 Hz, energy swing-up handing over to LQR, parameters identified on the rig, MATLAB app and Simulink model. |
| rhino-rmcs-2303-dc-servo-control | Speed and position control of a 100:1 geared encoder motor over Modbus, from an Arduino Uno. 201 runs across four days; the identified model predicts unseen runs to about 5 %, including a textbook PI design that limit-cycles on the real drive. |
| arduino-temperature-control-lab | Bulb heater, DHT sensor, solid-state relay. On-off through PID, and the firmware's safety behaviour — over-temperature cut-off, sensor-failure stop, host watchdog — tested against an emulated plant. |
| arduino-water-level-control-lab | Pumped tank with an HC-SR04 level sensor. Echo rejection, median filter, a rate gate on how fast water can physically move, dry-run and overflow guards. |
| pneumatic-circuits-openmodelica | ISO 6358 valves and cylinders written from first principles in OpenModelica. Two-hand press, reciprocation and pressure-stamping circuits, cross-checked against an independent Python model. |
| rigid-body-simulator | 3D rigid-body dynamics in MATLAB from Newton–Euler up: a thrown die whose landing face is read off the simulated orientation, a 3R arm with recursive Newton–Euler dynamics, a four-bar linkage solved through loop closure. Interactive app, 118 tests, CI on Linux, Windows and macOS. |
Alpino — decision-support tooling for a health foods business: OCR invoice reconciliation, attrition risk scoring, RFM segmentation, behind a Streamlit front end. Python, not control.
csc210-csc530-assessment (private) — self-hosted coding-quiz platform on Moodle, CodeRunner and Jobe in Docker, for running programming assessments at course scale.
How I work: conventions, frames and units fixed in writing before any code; results checked against a closed form, a conservation law or a second derivation; the number kept rather than the impression; and one entry point plus CI, so it runs on someone else's machine too.
Tools — MATLAB and Simulink · C/C++ on STM32 and Arduino · ROS 2, MoveIt 2, ros2_control, Gazebo · Python · Modelica · Docker · Git and GitHub Actions
Industrial robotics (from the KR3 platform) — KUKA KR C4 and RSI over UDP, with a protocol stack in C and a controller simulator to test it against · ros2_control hardware interfaces and trajectory tracking · MoveIt 2 with Pilz industrial motions (PTP, LIN, CIRC) and Servo · Gazebo Fortress physics co-simulation · a cell supervisor state machine with motion gate and watchdog · fidelity levels from mock hardware through physics and protocol simulation to the real arm · rosbag2 MCAP, an SQLite experiment database and Foxglove for analysis · Docker with the NVIDIA container toolkit and resource limits · gtest end-to-end launch tests, sanitizers and CI