Machine vision based autonomous forklift
ANSI B56.1 and B56.5 compliant (currently limited by software scope to unloading trailers)
3000 lb lift capacity (unchanged from unmodified Mitsubishi FBCS14N)
Gross weight approx 9000 lb
Added Lidar sensor suite as well as Realsense 3D camera
Onboard PC for path planning and ML models
PLC layer interfaces via CAN with factory Curtis motor drivers and sensors
For the developmental phase of the project I was the only ME on staff, and I designed the entirety of the mechanical integration hardware personally.
The major challenge with this project was keeping up with all of the evolving requirements. A lot of hardware ended up changing frequently, and the model shown here is actually the second chassis Fox adapted. The original was a Jungheinrich ETG235 model that was originally put into production in the 90s. The models we had were produced in the early oughts. Their controllers weren't particularly friendly to integration efforts, so the electronics were completely gutted, with only the induction motors, their onboard encoders, and hydraulics remaining.
The production version shown here is a Mitsubishi produced version of the Jungheinrich ETG214 (FBCS14n) that comes direct from the factory with Curtis controllers. They can be flashed with a software image that Fox owns, so the rest of the upgrade is essentially bolt-on kits of sensors, power electronics, and computational assets.
Only pictures shown are shown here to avoid infringing on IP, with the exception of some publicly available COTS part models.
The early phase of the company was entirely developmental, and was a fantastic experience. It took about a quarter to get the system physically operational, and then about a year to make it reliable and figure out the intricacies of the use case and the operational details. Initially the plan was to obtain the same base forklift model as the prototype for a production series, but once the robustness and capability of the machine vision system and behaviors were out there, it was possible to get factory stock of a new model. That meant a complete redesign of the physical hardware.
Sick's products are great, but some of their specs are a little conservative, and they recommend a clearance window of 2in centered on the laser's plane of rotation. That's a pretty generous space to let random warehouse debris impact the polycarbonate and cause either wear or impact damage. Sick has a really nice obscurant detection system, but that doesn't eliminate the need for repairs. Pallet stringers can be as thin as .5in, so it was easily worth the trouble to identify the functionally minimum window. Also, some cheaper lidar had in the past had some issues with runout in their rotating mirrors, turning their perception manifold from a plane into a cone. While the MS3 was never a problem in this regard, it wasn't much effort to set up a system to evaluate that too on other systems in case we wanted to try them.
You can see the laser in this video. The "red" glowing lights in the upper right are IR LEDs that shine from the body, through the window, and off a retroreflective target in the furthest part of the housing, which lets the system detect dust and other occlusions. The light pink strip that moves up and down with the camera is the laser. One of the fun things about working with nice equipment like this is getting acquainted with the great work people have done to flesh it out and make it a solid system.