Robots that work on your desk often fail because they are trained only on perfect examples, can't correct mistakes, and lose track of the big picture.
The system is trained on millions of images, text instructions, and robot moves from many different robot designs. It uses a method that saves both its own error and a human's correction, so it learns from mistakes. It also keeps an event memory to remember the overall task.
On a standard test for desk robots, it succeeded 93.6% of the time. On a difficult wire-insertion task, it reached 70% success. This shows that large-scale data, error correction, and task memory can make desk robots far more reliable.