Our vision for the future of agriculture includes the use of robotic systems that continuously work on the fields. They will monitor plant growth and health, weed and bring out seeds or fertilizer. In order to accomplish these tasks, the robots need to make goal-directed decisions, based on knowledge and facilitated by employing high-level reasoning and planning systems. Furthermore, their deployment is not limited to a single mission. Instead, long-term autonomous systems need to take care of a multitude of tasks without continuous human supervision.
These requirements lead to specific problems that must be addressed:
1. A high degree of autonomy requires a very feature-rich and therefore complex robot control system. Currently, most of such systems are rather monolithic, static software configurations, which are hard to maintain, upgrade, and reproduce.
2. Safety and reliability constraints become highly relevant as the robots will work unsupervised most of the time. Therefore, it must be assured that harm to humans, wildlife or the environment is excluded.
3. Behaviour constraints must be fulfilled. They are based on process-relevant knowledge like weather, soil conditions, type of seed, ownership or regional laws enforcing strong documentation duties or bans for activities in certain areas.
4. Self-preservation of the robot must be assured. Regardless of the high-level planning system used, the robot must be able to detect critical system states. A low battery, sudden changes in the weather or other events must be responded to accordingly by reorganizing the current task, aborting or pausing the current mission, or switching to another adequate recovery mechanism.
In this work, we propose concepts to apply cloud computing and containerization methods described in previous work to knowledge driven, long-term autonomous robotic systems in the agricultural domain. These concepts allow a great increase in terms of modularity, scalability and thus maintainability and reliability of the deployed robots. Furthermore, we present a concept to include multiple safety layers to the knowledge-driven architecture of our prototype robotic system that can trigger re-planning or override the previously planned actions of the corresponding higher level.
This research project was presented in the scientific seminar organised by Robagri during the 2019 FIRA:
Presented by M. Höllmann, German Research center for Artificial Intelligence.