VEXAutoSort
From our own bins to every team's bench.
How VEXAutoSort goes from a fix for our own sorting headaches to something any VEX team can build and run.
The spark
The idea hit during manual sorting at the World Championships. Every team in the room was losing the same hours to the same boring job, so we figured it was worth solving once, for everyone.
A nudge from the pros
A call with Shawn Patel, CTO of Almond Robotics, pointed us toward the SO-ARM101 and an approach where the robot learns by copying a human. He also flagged the hard part: staying reliable across 50-plus different parts.
The mechanical breakthrough
The insight that makes this buildable: don't make the robot classify every part. Pair it with simple 3D-printed sorters so the arm only learns one motion, which keeps the design cheap and easy to reproduce.
Hardware online
Got the arm initialized, calibrated, and running in teleoperation, then added the first camera.
Cameras, lights, and a gripper
Mounted an overhead and a wrist camera, printed the mounts, and added lights so the arm gets a clean, consistent view. All of it off a normal parts list.
First demonstrations
Started with one move: pick a gear and drop it in a bin. We recorded about 200 teleoperation demos and got pickier about data quality each run.
First working pick-and-place
Trained an ACT policy and got the pick-and-place running on its own at around 90%. The big fix: the arm kept blocking the overhead camera as it reached in, so the wrist camera let it see past its own gripper.
From one part to a full bin
Sort two different parts into their own bins, then scale to five or more, tuning cameras and lighting as new mix-ups show up.
Lock the reference design
Freeze the hardware layout, parts list, and configuration into one known-good setup that a second team can build from.
Docs anyone can follow
Write step-by-step build and setup guides, record clear demo videos, and clean up the codebase so a team with no ML background can get sorting.
Bill of materials and cost
Cost out every component and aim for a price a high school team can actually afford.
Pilot with real teams
Get one or two other VEX teams to build VEXAutoSort using only our docs. Wherever they get stuck is what we fix next.
Decide how to ship
Open-source the design, sell ready-to-run kits, or both. The goal: any team that wants to stop sorting by hand can.
A sorting robot on every team's bench.
We started this to fix our own problem. We're finishing it so no VEX team ever has to sort a bin of screws by hand again.