Unmanned Systems Technology 038 l Skyeton Raybird-3 l Data storage l Sea-Kit X-Class USV l USVs insight l Spectronik PEM fuel cells l Blue White Robotics UVIO l Antennas l AUVSI Xponential Virtual 2021 report

74 trialling all the vehicles and systems in the field to make sure they work correctly before deployment.” Local maps are uploaded to the Platform, through which Blue White’s AI algorithms plot standard paths for the different tractors and attachments based on their jobs. They also reconfigure those paths on an ad hoc basis to prevent traffic jams between farming vehicles. Fine adds, “We are building a short on-site curriculum plan for training the grower, if they want to handle the operation themselves or hand it to their own people. It encompasses the use of the control station and interface as well as operational safety training.” A typical mission Once all the necessary subsystems have been installed, only one operator is needed to operate a typical tractor fleet of up to 12 vehicles from the user interface. From here, field tasks can be selected from a list, and for each task the appropriate tractor will be deployed onto the correct route of waypoints into, through and out from a given field. If a Blue White operator is hired to operate the fleet (since not all farmers would wish to do so), they typically opt to do so on-site, for swift communication and feedback from the farm personnel, although they can also work remotely if the grower prefers. “As the attachments vary between tasks, it’s critical that we automate tractors to hold a really tight centreline down rows of crops or trees – even amid vibrations, shocks or drags that the trailer imparts – to minimise the chances of collisions,” Fine says. “Most of our r&d goes into making our perception system smarter and more accurate, via the multi-sensor solution,” Ascher adds. “We use sensors from different vendors, as we want to stay component-agnostic. We’re not looking for the most high-end sensors available, for example, just the best ones within our goal cost range.” Each tractor will run for upwards of 12 hours on a full tank of fuel, typically at a slow speed for most tasks such as mowing and spraying, with occasional stops if an attachment needs to be refilled. Between each row of crops, the driving algorithm takes perception and navigation inputs from the Lidar, cameras and GNSS-INS, and filters them through context-specific requirements such as orchards requiring tighter clearances (and thus more intensive computations) than trellises or hedges. Lidar and cameras – the perception system – typically form the primary line of safety in this regard, given that GNSS outages are frequent when tractors pass amid trees, and satellite navigation cannot pick up random obstacles such as animals, humans or other vehicles in front of the tractor. “We fuse all the obstacle avoidance data as well as any observed warning lights in the tractor in our high-level computer,” June/July 2021 | Unmanned Systems Technology At the ends of rows of crops, the tractors can update their GNSS feed to re-localise themselves, much as an AUV does after resurfacing