Studying mechatronics and control
with an open mind.

Vortex NTNU's AUV needed a manipulator arm that could be commanded precisely and safely to complement the vehicle's valve-manipulation capabilities. The subsystem also had to be built, reviewed, and handed off across a rotating student team.

On the control team, I designed the ROS 2 message and action interfaces in vortex-msgs, then built the reference-filter guidance module, controller skeleton, CAN interface, and Stonefish simulation interface in vortex-gripper.

The action server lets the wider autonomy stack drive the gripper through each command and phase of an autonomous mission, tracking or cancelling a goal mid-execution, with a threshold on state error triggering the final snap to setpoint. Stonefish verifies this logic end to end in simulation, while the CAN interface carries it through to the physical hardware.

Validated end to end in Stonefish simulation, with integration tests in vortex-auv covering full goal sends through the pipeline.

NTNU's TTK4145 Real-Time Systems project: a peer-to-peer network of elevator nodes has to keep serving hall and cab calls correctly through motor faults, door obstructions, and dropped network connections, without ever needing a manual restart. Each node broadcasts a state snapshot over UDP every 10ms and merges its peers' snapshots into a shared consensus view, so no central coordinator can become a single point of failure.

Built with a three-person team for the semester project. The system runs as five concurrent Go modules communicating over channels: a driver polling buttons, floor sensors, and door obstruction; a finite-state machine turning assigned orders into movement; a network node handling consensus; a manager assigning hall and cab requests; and the door and lights output.

Each hall and cab request carries its own cyclic counter (Unknown, Inactive, Requested, Active) rather than one counter for the whole state snapshot, so conflicting reports of the same order resolve locally without blocking on every other request. An earlier version counted the whole snapshot instead and was scrapped: it could only resolve one worldview at a time, which broke handling of concurrent requests.

TTK4130 Modelling and Simulation, assignment 5: derive the Newton-Euler equations of motion for a cube-shaped satellite and predict its torque-free attitude dynamics over one orbit. A uniform cube spins about a fixed axis indefinitely, but bolting a small mass onto one corner makes the same satellite's spin axis precess instead, the textbook difference between a spherical and an axisymmetric top.

Solo coursework: modelled the satellite as a 13-state system (position, velocity, attitude quaternion, body angular velocity) and derived the corner-mass case's shifted inertia tensor by hand via the parallel axis theorem. The orbit and torque-free rotation were then integrated numerically over a full 139-minute orbit.

Attitude is represented with a unit quaternion, avoiding the gimbal lock that Euler angles introduce, and the torque-free Euler equation is integrated directly in the body frame. Angular momentum and rotational energy stay conserved to around 1e-15 relative drift across the run, a hard numerical check on whether the integration is actually correct.

Angular momentum and rotational energy hold to machine precision, about 1e-15 relative drift, across the whole simulated orbit for both cases.

NTNU coursework, solo: nearest-neighbour and k-means classifiers built from scratch, to see the algorithms at the level of the underlying maths. 1NN on raw MNIST digits reaches 96.91% accuracy; compressing the reference set to k-means templates and running k=7 against them trades about 3.4 accuracy points for a far smaller reference set (93.49%), with errors concentrated in a few genuinely ambiguous pairs, like 4s read as 9s.

Self-directed, outside any coursework or team assignment: I found the public ArcGIS services behind Norwegian housing-price maps, scraped and modelled the data myself, and ran an IAAO-standard valuation pipeline end to end. The write-up reports the model's honest accuracy numbers alongside the wins.