andlyu/blupe-toys

blupe-toys is a collection of small physical setups, called toys, designed to let a robot arm practice a task over and over without a person resetting the scene each time. Testing a robot policy properly usually needs forty to fifty attempts per checkpoint to know with any confidence whether it is actually improving, and normally every one of those attempts needs a human to put objects back in place and judge whether the robot succeeded. That does not scale well when you are trying to improve a robot policy quickly. Each toy in this repository is built so that resetting it is easier than doing the task itself, and success can be checked automatically without a person watching. The repository includes three such toys. One has the robot drop a light blue ball into a tall black cylinder, where the cup is shaped so the ball naturally rolls back out to a random spot afterward. Another has the robot place an object on a small shelf that automatically flips over and dumps the object back off when checked. The third has the robot reach out and press a red button that is actually a USB device, so pressing it registers electronically rather than needing a camera to confirm. Because these evaluations reset and score themselves, they can run continuously while a person is away, doing something else, or asleep. The project states this approach helped push one robot policy, MolmoAct, from 62 percent success up to 91 percent over five iterations, using around 30 minutes of periodic attention instead of 5 hours of a person manually resetting the scene. The toys were originally built around the SO-101 robot arm, but the setups are not specific to that arm and should work with any arm able to reach a small tabletop area. Each toy's folder includes the parts list and setup instructions needed to build it. The project is written in Python and currently has 2 stars on GitHub.