Programming two robot hands to coordinate complex movements like picking up, adjusting, and passing objects is very hard. Engineers usually have to write special instructions for each action and each task, which is time-consuming and can't be reused.
The researchers trained a robot using a single simple rule: move the object to a target pose. Instead of tailoring rewards for each behavior, they varied the training scenarios. The robot's two arms discovered for themselves how to pick up, adjust grips, and hand over objects to reach the goals.
The trained single policy succeeded in 61 of 80 real-world trials (76.3%) on four different tasks with objects it had never seen before, requiring no task-specific adjustments. This suggests a straightforward way to build general two-handed dexterity by just specifying what should happen to the object.