Evaluating the PORTABILITY of the FROBOMIND ROBOT SOFTWARE architecture to new AUTONOMOUS PLATFORM

To be able to use the same software solution on their Platforms University of Hohenheim set out to implement FroboMind on the Autonomous Mechanisation System (AMS). FroboMind was ported and implemented on the AMS within the time limit of five days. The University of Hohenheim (UniHoh) is conducting research in novel precision agriculture production methods which involve crop scouting and organic weeding applications etc. and are collaborating with the University of Southern Denmark on utilizing autonomous field robots in these projects. UniHoh has a field robot platform The AMS which is based on a Hako tractor retrofitted with sensors, actuators and the MobotWare framework (Reske-Nielsen et. al., 2006), (Griepentrog et al., 2011), (Griepentrog et al., 2009) allowing autonomous navigation in the fielda and orchards. Recently UniHoh acquired a new robot platform called Armadillo Scout which is a modular platform designed for precision agriculture research. Armadillo Scout runs MobotWare (Beck et. al., 2010)
and the open source FroboMind architecture (Jensen et. al., 2012) based on Robot Operating System (ROS) (Quigley et al., 2009). MobotWare has successfully been ported to both the AMS and the Armadillo Scout. To be able to use the same software solution on both platforms UniHoh set out to implement FroboMind on the AMS in this project. The purpose is to have FroboMind and Mobotware coexisting on both platforms which makes it possible to choose the best suitable solution for a given research project based on the qualities of Mobotware and FroboMind. The aim of this work is to evaluate the portability of FroboMind to a new robot platform with respect to both the implementation process and performance of the robot when operating autonomously in a precision agriculture environment. In order to test the FroboMind portability with respect to reliability and precision agriculture applications it was decided to perform a field test in an apple tree orchard at UniHoh. The tree rows were 100 meter long and interspaced by 4 meter. In some rows a few trees were missing and/or replaced by younger and smaller trees. In the first trial the AMS mission was to navigate autonomously between two rows of trees in an apple tree orchard. At the end of the row the AMS would make a 90 degree left turn, drives straight, and make a 90 degree left turn which would position the robot in an adjacent row. It was decided to repeat this process continuously 12 times corresponding to 6 full rounds driving the same track. In the second trial the AMS mission was to navigate autonomously through the same apple tree orchard alternating between left and right turns. Before the trial began it was decided that the focus would not be on the sensor/controller performance under difficult circumstances, so two persons stepped in as "tree" whenever one of the rows had a large hole between the trees.
During the first trial the following observations and comments were made: The first round was completed successfully. The second round was completed successfully. During the second round a video recording was made while walking behind the robot to document the orchard environment. During the third round the robot navigation failed one time causing the robot to drive into the trees. This happened at a location where the apple trees were replanted by much younger trees at the left row. During the fourth round the robot navigation failed one time causing the robot to drive into the trees. This happened at another location where apple trees were missing for more than 6 meters at the right row. The fifth round was completed successfully. During the sixth round the robot navigation failed one time causing the robot to drive into the trees. This happened at the same location was the fail during the third round. The navigation failures occurred at locations where the apple trees did not provide sufficient data for the implemented laser range scanner row detecting algorithm. Aside from the described problems the robot navigated reliably through the orchard. The second trial was completed without any errors in the navigation.
For a 4 man team it is possible to implement and test FroboMind with 5 working days. The implementation is only possible as long as the hardware interfaces are well documented and the platform is working. It is not possible to optimize the behaviour algorithms or low level control algorithms nor is it possible to debug and troubleshoot small problems on the platform within the time frame.