Anytime ROS 2: Timely Task Completion in Non-Preemptive Robotic Systems

Ensuring timely and predictable data propagation is essential in real-time robotic systems, yet they increasingly incorporate computationally expensive, long-running tasks, such as deep neural networks. Integrating these tasks into the Robot Operating System 2 (ROS 2) is challenging due to the non-preemptive design of the ROS 2 executor, which can lead to high task interference and unpredictability. This paper presents a structured approach to transform these long-running tasks into a series of smaller, more manageable subtasks. The transformation inherently introduces preemption points, which enable an anytime functionality where tasks can be terminated early while delivering progressively better results. The approach leverages the existing ROS 2 actions framework to provide a reliable task-cancellation mechanism and feedback channel. This solution requires no modification to the existing ROS 2 codebase and remains compatible with existing ROS 2 end-to-end latency analyses. The evaluation of Anytime RRT* and AnytimeYOLO demonstrates that timely results can be achieved via flexible cancellations.