The Need for Robot Learning
Why robots need to learn, and how they are learning to work outside the factory.
Why robots need to learn, and how they are learning to work outside the factory.
Robots Outside the Factory
Industrial robots perform repetitive tasks with a speed and accuracy far beyond ours. Yet most of them still work in carefully engineered settings: fenced cells on factory floors, where every part arrives at a known place, in a known orientation, at a known time. These robots execute programs written for one task and one environment. If the task or the environment changes, an engineer has to reprogram them. A car-assembly robot that picks up wheels from a feeder and bolts them to an axle relies on the feeder being exactly where its program expects it to be. Move the feeder, or ask the robot to remove wheels instead of mounting them, and the robot is of no use until someone rewrites its program.
Why Programming Is Not Enough
Homes, hospitals, farms, construction sites, and most workshops are very different. Every room has a different layout, objects come in endless shapes, sizes, materials, and weights, and things constantly move around. Programming a robot in advance for every object, every layout, and every combination of task and tool is impossible. A robot that is to help in these places must instead learn: from demonstrations by people, from its own trial and error, and from experience accumulated across many tasks, so that it can adapt to situations that nobody anticipated when it was built.
How Robots Learn Today
Robot learning has progressed rapidly in recent years, building on the same advances in machine learning that produced today’s language and image models. A few ideas play a central role:
- Learning from demonstrations. A person shows the robot how to perform a task, for instance by guiding it remotely, and the robot learns a policy: a mapping from what it sees and feels to what it should do next.
- Learning from experience. Through reinforcement learning, a robot improves by trying, observing the outcome of its actions, and favoring the actions that work. Much of this practice can happen in simulation, before the robot touches the real world.
- World models. A robot can learn to predict how the world will respond to its actions, and use these predictions to plan, much as we imagine the consequences of a movement before making it.
- Large pretrained models. Models trained on vast amounts of images, video, and text give robots a broad understanding of objects and instructions, which they can then refine for physical tasks. This combination of AI with a body that acts in the physical world is often called physical AI.
Examples from My Group
In my own research, robots have learned where to place their fingers to grasp objects they had never seen, by recognizing parts (handles, rims, tops) by which similar objects were grasped before, and transferring this knowledge to the new objects. They have learned to feel whether a grasp is stable, using tactile sensors to detect that an object is about to slip before lifting it. More recently, my group has developed robots that assist people with paralysis by anticipating where they intend to move a robot arm, robots that spray and finish concrete on construction sites despite dust and poor visibility, and learning methods that let robots plan and act over long sequences of steps. See the Research page for details.
In all of these cases, the “program” that governs the robot’s behavior is no longer written once and for all: it keeps evolving as the robot gains experience.