Robot learning is the approach of creating robot behavior by learning from data and experience rather than by hand-writing a controller for each task. Classical robotics engineered explicit models of the robot and its environment and derived control laws from them - precise and reliable in structured settings, but brittle in the unstructured messiness of homes, warehouses, and roads. Robot learning instead trains policies that map sensor inputs to actions, drawing on the same machine-learning toolbox that transformed vision and language.
It spans several techniques. Reinforcement learning lets a robot improve by trial and error against a reward, often in simulation and then transferred to reality via sim-to-real methods. Imitation learning, and its simplest variant behavior cloning, teaches skills from human demonstrations. Since roughly 2022 the field has converged on a foundation-model framing: train one large model on broad, diverse data and adapt it to many tasks. Google’s RT-1 demonstrated this with 130,000 demonstrations across 700 tasks, the Open X-Embodiment collaboration pooled data from 22 robot types to show experience transfers across embodiments, and vision-language-action models inherited web-scale knowledge to generalize to new objects and instructions.
Why business readers should care: robot learning reframes robotics from a per-task engineering problem into a data and model problem. If general robot foundation models keep improving the way language models did, deploying a robot to a new task could become a matter of fine-tuning and demonstration rather than months of bespoke programming.