Cup Stacking Notes
Notes on teaching a low-cost robot arm to stack paper cups.
11 posts, in the order it happened. Each one ends with the open questions.
Overview · Oct 2026The setup and the goal
A six-joint arm, two cameras, two paper cups, and three ways to teach it to stack.
2 min read
Before Oct 21The first ACT model, and a robot that would not move
20 demos, a clean training curve, and an arm that stayed at rest.
2 min read
Oct 21 to 22Trimming idle frames, and every way a camera can fail
A cleaning tool, a retrain, and a list of things that go wrong between a good checkpoint and a moving arm.
3 min read
Oct 22 to 23Going to 101 demonstrations
More data, one instruction, and a trimming job that was far too slow the first time.
2 min read
Oct 23 to 24Fine-tuning SmolVLA on a rented GPU
Batch 64, 10,000 steps, a memory crash at step 800, and about $4.80.
2 min read
Oct 23Skipping the neural network: a camera-guided stacking skill
Computer vision, inverse kinematics, a wrist-roll cap that was wrong, and centering the cup with two cameras and no markers.
4 min read
Oct 24Why SmolVLA looked bad on the robot
A loss of 0.013, an arm that barely moved, and two mistakes that were mine.
3 min read
Oct 24, afternoonThe fix was the room light
SmolVLA started working when I switched on a lamp. What changed in the pictures, and why my offline test missed it.
4 min read
Oct 24, eveningDoes the model look for cup B?
With cup B moved far away, SmolVLA carried cup A to where cup B used to be. Why, and what data would fix it.
3 min read
Oct 24, nightACT on the same data: the same story
The small ACT model stacks cups from the usual positions and fails when I shift them, just like SmolVLA.
2 min read
SummaryEvery run in one table, and the lessons
What was tried, what each run cost, and what I would tell myself before starting.
3 min read