This master thesis investigates the design and evaluation of a deterministic, real-time control architecture for robotic manipulators using the ROS2 ecosystem. The central challenge addressed is achieving hard real-time guarantees in a middleware environment that was originally designed for soft real-time performance.
The architecture leverages the ros2_control framework in combination with a patched real-time Linux kernel (PREEMPT_RT). A custom hardware interface plugin was developed to communicate with the motor drives over EtherCAT at 1 kHz, while a separate real-time executor schedules control-loop callbacks with bounded jitter below 200 µs.
A comparative evaluation was performed against a non-real-time baseline, measuring jitter, latency histograms, and trajectory tracking error under various CPU load conditions. The real-time configuration reduced worst-case jitter by 94% and improved trajectory tracking RMS error by 37% during high-load scenarios.
The thesis also proposes a component isolation pattern that wraps computationally expensive planning nodes in dedicated processes with relaxed scheduling policies, decoupling them from the time-critical inner control loop. This pattern was validated on a 7-DoF KUKA iiwa arm and is now being considered for integration into upstream ros2_control.
The full thesis document, source code, and benchmark datasets are available on GitHub. The work received a grade of 1.0 (with distinction).