TRASHCAN-CV_

We wanted a demonstrator that forced perception and motion to agree in real time, on a machine that fails visibly when it gets it wrong. A bin that catches thrown rubbish is unforgiving in a useful way: the object is in the air for about 700 ms, the answer has to be right the first time, and everyone in the room can see whether it worked.

The internal brief was one sentence and one number: catch a thrown object more often than not, from anywhere in a 4 × 4 m room, in under six weeks.

Two direct-drive hub motors, differential steering, and no casters — a caster at 2.4 m/s with a hard stop puts the whole mass into a shimmy we could not tune out. Instead the bin is a statically unstable two-wheeler held upright by the controller, which cost us a week of firmware and bought back 300 g and about 40 ms of settling time.

The lid is the part we got wrong twice. The first version was a servo-driven flap that had to be open before the object arrived, which meant committing to a catch 200 ms earlier than we wanted. The second used a torsion-spring lid released by a solenoid: 60 ms to fully open, no commitment until the last moment. That is the version in the drawing.

One 4-layer board, 82 × 96 mm, carrying a dual field-oriented-control stage and the power tree. We split vision and control deliberately: an Orin Nano runs the detector, an STM32G4 closes the motion loop at 20 kHz, and they talk over CAN at 1 Mbit.

Putting the loop on the same processor as the model was tried and abandoned in week two — inference jitter was showing up as steering wobble.

Two global-shutter mono cameras at 120 fps, 1440 × 1080, baseline 180 mm, hard-synced. Rolling shutter was never an option: at 8 m/s a rolling sensor smears the object across enough rows to move the centroid by more than our capture window.

Detection is a small single-class detector trained on 11,000 frames we collected ourselves over two days with a throwing rig, quantised to INT8 and running at 120 fps with 9 ms of headroom. It only has to find one thing: an object in flight. Everything else — trajectory, impact point, whether to commit — is a ballistic fit over the first 90 ms with drag ignored, which is wrong in principle and accurate to ±80 mm in the room we care about.

250 recorded throws by nine people who were told to try to beat it. 210 caught. Of the 40 misses, 23 were the backlit-window case, 9 were objects with lift, and 8 were genuinely bad predictions.

It ran 100 throws unattended with three misses, which is the run we consider the deliverable. The traverse-and-settle target was missed by 30 ms and we left it missed — closing it means either more torque or less mass, and both were the wrong trade for a demonstrator.