Pastures constitute the largest land-use category in the United States. According to the USDA Natural Resources Conservation Service, 27% of the total acreage of the contiguous 48 states is used as privately-owned ranges and pasture lands. These lands are often actively managed. In particular, they are mowed routinely to prevent diseases or non-nutritious weeds from taking over. Managing these vast lands, often on rugged terrain, is a difficult and laborious task. An ideal job for a robot!
The Robotic Sensor Networks (RSN) Laboratory, led by Professor Volkan Isler, in collaboration with CFANS’ West Central Research and Outreach Center (WCROC) located in Morris, MN and the Toro Company has been collaborating on building the CowBot, a robotic mower capable of autonomously operating on cow pastures. The project is sponsored by the Legislative-Citizen Commission on Minnesota Resources (LCCMR) program. CowBot is built on the Toro Groundmaster platform that was modified by Toro to operate with a 28.8 kWh electric motor powered by a lithium battery pack. CowBot incorporates a suite of safety features including a wireless remote emergency stop button, pressure activated perimeter bumpers and emergency stop knobs on the front and back of the mower. It uses a flail-deck weeding implement custom mounted with high ground clearance to achieve a cut-height of 20 cm (8 in.). The high cut ensures sufficient volume of grass for the cows to graze while also stunting weed growth. CowBot is supplemented by a solar-powered mobile recharging station designed and developed as proof of concept by the WCROC researchers. This way, CowBot can recharge on the field without the need for any fixed infrastructure.
RSN researchers have been developing the autonomy stack of CowBot which includes both low-level navigation capabilities and high-level coverage planning. At the moment, CowBot uses GPS, inertial sensors and odometry to accurately follow a given trajectory. The team has also developed efficient mowing path planners and demonstrated them in successful field trials. CowBot was a big hit at Minnesota Farm Fest 2021 where it provided live demonstrations.
The development of the CowBot platform led to interesting research problems as well. For example, if the locations of the weeds can be determined in advance, the CowBot can target them directly (instead of mowing the entire field.) Aerial imagery obtained from off-the-shelf aerial drones can be used to obtain the weed map: Once the images are obtained, a classification algorithm can detect the weeds whose locations are passed on to a planning algorithm to determine CowBot’s trajectory. The performance of the weed classification algorithm clearly affects CowBot’s performance: One can choose a low error threshold to ensure that all weeds are detected. But then, resulting false positives lead to longer paths for the robot. If one chooses a higher threshold to reduce false positives, too many weeds can go undetected. Post-doctoral researcher Parikshit Maini and Professor Isler have recently established a mathematical relationship between the classifier performance and the coverage trajectory. Their results provide a systematic way to address the trade-off between detection accuracy and the length of the robot trajectory. The team, along with PhD student Burak Gonultas is now working on removing the need for aerial imagery. Instead, they are equipping the CowBot with onboard cameras and LIDAR so that the weeds can be targeted directly using the information from these sensors. This is a difficult problem: due to the limited footprint of these sensors, the CowBot does not get to see the entire field and therefore must plan its trajectory using only partial information.
Another interesting research problem addressed by the team is regarding energy efficiency. As part of his PhD thesis, Minghan Wei worked on computing coverage strategies which incorporate CowBot’s energy limitations. In particular, in order to cover a large field, CowBot may have to return to a base station and recharge multiple times. Wei and Isler developed coverage algorithms with theoretical performance guarantees to ensure that the field is mowed in a close-to-optimal amount of time.
The CowBot project is now in its final year but there are many interesting research and development opportunities for future work. In particular, RSN Lab researchers are interested in developing dynamic obstacle detection and avoidance algorithms to make CowBot completely safe. They also think that CowBot can be used for herding the cows and become one of the best friends of ranchers!