Pi robot runs box assembly at Dandelion Chocolate
Chelsea Finn says the hard part was not the dexterous box folding, it was putting the finished boxes on the pile without knocking it over.
Chelsea Finn posted a video of a Pi robot deployed at Dandelion Chocolate building boxes, and says it is fully autonomous with no interventions. The deployment recently moved from a table-mounted robot to a mobile one, and she says the robot is now completely self sufficient for multiple hours.
Counterintuitively, the hardest part was reliably stacking the boxes.
The interesting part is which step broke first. Finn says she expected box assembly to be the hard problem, because folding a box is the most dexterous thing in the task. It was not. The hardest part was reliably stacking the finished boxes.
Why stacking is worse than folding
Three reasons, by Finn's account. The original table-mounted robot had poor visibility of the stack without separately mounted cameras. Stacking also demands more generalization, because each box goes in a different spot rather than the same one every time. And errors compound. An imprecisely placed box can bring the whole stack down many boxes later, long after the mistake happened.
That last point is what researchers picked up on. Chris Paxton summarized it as stacking being harder than making a box because each step has to be near perfect or the pile collapses. It is a failure mode that does not show up when you evaluate a policy on a single box in a lab, which is exactly Finn's framing. She describes the lesson as the gap between proof of concept, building one box, and real world utility, productively building boxes for hours.
Shown versus claimed
The no interventions and multiple hours claims come from Finn, and nobody outside Pi has verified them. The video shows the robot working at a named business rather than a lab bench, which is more than most manipulation demos offer, but no throughput or cycle time numbers were posted. So how the robot compares to a person doing the same job is unknown.
Finn also framed the deployment as a data source rather than just a demo, saying feedback from real world use cases like this feeds the Pi pre-trained model as the company scales. That is the quiet argument here. The value of putting a robot in a chocolate factory may be less about the boxes and more about learning which step actually fails.
Stacking was harder than making a box because each step must be completed near-perfectly or the pile will collapse.