It may soon be possible to teach a robot any task just by showing it how it's done -- a single time.
Researchers at UC Berkeley have developed a way to speed up the education of our silicone-brained friends. In a recently published paper, they report on a new learning algorithm that allows robots to mimic an activity it observed just once on video.
Training robots today is hard work. Even really simple actions like picking up a cup require paragraphs upon paragraphs of code expressly telling the bot what to do each and every step of the way -- a process that is hard, complicated, and probably frustrating for us humans.
There's work to do even after the code is fully laid out. For example, take assembly line workers. After all the instructions are copy-pasted into their circuits, these bots must undergo a long training process during which they must execute every procedure multiple times. They do so until they can perform the task without making any mistake along the way.
More recently, programmers have created software that allows robots to be programmed just by observing certain tasks. While this is more similar to how we or an animal would learn, it's still clunky to use -- currently, we need to show our robotic friends such training videos thousands of times until they get the hang of it.
The team from UC Berkeley, however, describes a new technique they developed that allows robots to learn a certain action just by observing a human do it a single time.
This technique combines imitation learning with a meta-learning algorithm, the team reports. They christened the resulting system 'model-agnostic meta-learning' (MAML). Meta-learning basically means 'learning to learn'. MAML is a process by which a robot builds on prior experience in order to learn something new. If a robot is shown footage of a human picking up an apple and putting it into a cup, for example, it can gauge what its objective is -- putting the apple in the cup. As it learns how to handle these objects, it can expand that knowledge to other similar behaviors. So, for example, if you then go on to show it a video of somebody putting an orange down on a plate, it can recognize the overarching behavior and quickly translates that into the motions it needs to do to carry out the task.
Best of all for all those assembly-line robot trainers out there, the bot doesn't need to know what an orange or a plate is -- it will still perform the required task.
In short, MAML provides a platform that allows a neural network (or a robot) to learn a wide variety of tasks starting with relatively little data. It's almost the polar opposite of how neural networks work today -- which master a single task while drawing on a huge dataset.
The team tested MAML on several robots. After a "single video demonstration from a human", they note, the robots could successfully perform the shown task. "After meta-learning, the robot can learn to place, push, and pick-and-place new objects using just one video of a human performing the manipulation," they conclude.
The paper "One-Shot Imitation from Observing Humans via Domain-Adaptive Meta-Learning" has been published in the pre-print journal arXiv.