As my stint as CTO of YellowLabel came to a close I was lucky enough to make an encounter that would revive my hopes of building autonomous vehicles for golf courses.
Despite my unsuccessful first attempt of converting mowers to be autonomous, I came to an insightful conclusion on the only possible way a small startup could penetrate this highly competitive market. The competitors were the likes of Toro and John Deere, both vehicle manufacturers with massive economies of scale, so the only change of bothering this duopoly was to develop and release an autonomous vehicle of much smaller scale than first envisioned. The conclusion was to build a bunker raking robot.
The first reason was simple, raking bunkers is the most man-hour intensive job in golf course maintenance. The second, and more key insight, was that we could do the same job in an unmanned vehicle significantly smaller than its non-autonomous counterpart – this would simply not be the case if we were to build a mowing machine. The size of its cutting blade apparatus is directly correlated to how fast it can complete the job as more grass surface area is covered in one mowing line. This thought process meant drastically decreased production and development costs, but they still were not low enough to justify the capital investment of building r&d facilities.
One night at a dinner party I was lucky enough to meet Alessandro Deodati, the CEO of Dronyx, a robotics company specializing in tracked robots for beach cleaning headquartered 30 minutes away from me! As they say, the rest is history. Within two weeks I struck a partnership with Dronyx, whom would be in charge of providing the adequate hardware, whilst I would continue to perfect and refactor our path-planning software for the new specifications.
At first I was convinced that it would be fairly straight-forward to apply the algorithm to the new robotic base by recalibrating for new encoders and vehicle dimensions, but this was an assumption that rapidly faded away. I ended up basically rewriting all the code from scratch since the bulk of the work was getting all the various sensors to talk to the path planning algorithm in the best way possible. It meant that the entire ROS node structure would be revolutionised, as changing just the way one sensor interacted impeded the functionality of others. The main challenges were dealing with a new type of GPS and IMU positioning input, reinterpreting signals from the motor encoders, and rewriting the stereo-camera algorithm to account for bunker edges instead of fairway edges. The hardest was definitely the positing inputs because apart from rerouting the data within the nodes in the most efficient way, I had to recode all the filters to normalize this data from the noise and combine it all into one single odometry node that continuously talks to the localization node (which knows exactly where the machine is on the course). The noise created by operating a vehicle of this type vs the original diesel tractor we used was completely different given the disparity in vehicle steps – vibrations from the engine were replaced by other concerns such as the bumps in the grassy terrain. After months of mathematical equations and on-field testing I was able to finally integrate the navigation software into the sleek new hardware projected and built from the ground up!
Our partnership was able to create a fully-functioning prototype!! Our estimations would put the production costs at around $8k, with a projected retail price of $20k, which is around \$5k less than its non-autonomous counterpart! Now the next challenge was to find the right strategy and channels to be able to commercialize the product in the best way.
Hardware projecting and proprietary software have been sold to a company based in Italy.