Automatic Parking — 4WS Vehicle

Conventional autonomous parking research focuses almost entirely on Ackermann-steered passenger cars. Four-wheel steering (4WS) platforms have a fundamentally different kinematic envelope — they can crab sideways, pivot on the spot and cover parking manoeuvres that a two-wheel-steered car physically cannot — but there was very little open-source work showing how to drive one through the full autonomy stack.

For my final-year Mechatronics thesis at Stellenbosch University I set out to close that gap: model a 4WS vehicle, simulate it inside the ROS2 / Gazebo stack, and make it autonomously park into a target spot using standard robotics building blocks (SLAM, a costmap, a path planner and a path-following controller).

I built the project bottom-up in ROS2 so that every layer could be swapped or tuned in isolation. Starting from the vehicle itself — forward and inverse kinematics for a 4WS platform — then a URDF and Gazebo model, then the perception and planning stack on top.

The autonomy stack is standard ROS2: LiDAR-based SLAM to build and localise against a map of the parking environment, AMCL for on-line localisation, A* as the global planner, NAV2's Regulated Pure Pursuit (RPP) controller for path following, and a mission layer that sequences approach, manoeuvre and final alignment. 4WS-specific behaviours — crab steering, in-place rotation — are exposed as primitives that the mission layer can call when the standard path doesn't fit the envelope.

Modelled the 4WS vehicle with both forward kinematics (wheel angles + speeds → body velocity) and inverse kinematics (desired body twist → per-wheel commands), with the two front and two rear steering angles treated as independent inputs. The model supports the three canonical 4WS modes: conventional, crab and counter-steer.

The ROS2 graph runs sensor ingest (2D LiDAR, odometry), a slam_toolbox node producing a 2D occupancy grid, AMCL against that map, the NAV2 stack (costmap_2d, A* planner, RPP controller) and a custom mission node that picks between standard NAV2 goals and 4WS primitives based on the target pose relative to the vehicle. Gazebo provides the physics, sensors and world; RViz visualises the live plan, costmap, and trajectory.

Code and launch files are open-sourced on GitHub with a minimal reproduction path — clone, build the workspace, and a single launch file brings the simulator, map, localisation and planner up together.

The thesis was awarded Cum Laude. The final system parks a simulated 4WS vehicle into pre-defined parking spots using a mix of NAV2-driven trajectories and 4WS-native manoeuvres, with live visualisation of the plan, costmap and vehicle pose in RViz.

Just as importantly, it became the project where I got fluent in the ROS2 / NAV2 stack — which directly set up the production robotics work I took on straight afterwards at BATTALION Technologies. A live TypeScript re-implementation of the 4WS kinematics and four scenarios — one per primitive — is available as an interactive demo, with the repo itself now polished to senior-engineer shape (CI, unit tests, docs, Docker).