Robot Learning

Using machine learning to let robots acquire skills from data and interaction, instead of hand-written rules.

Robot learning is the field at the intersection of machine learning and robotics: it studies how robots can acquire new skills or adapt to their environment through learning algorithms, rather than having engineers hand-code a control program line by line. The main approaches are imitation learning, learning from human demonstrations such as data collected via teleoperation; reinforcement learning, trial and error guided by reward in simulation or on real hardware; and, more recently, robot foundation models pretrained on large-scale data. The core challenges are that real-robot data is scarce and expensive, real-robot trial and error is risky, and there is a gap between simulation and reality. The field's dedicated venue is CoRL, the Conference on Robot Learning, and most model research in embodied AI falls under this umbrella.

ExampleCollecting a few dozen towel-folding demonstrations via teleoperation and training an ACT policy on them lets a bimanual robot learn to fold towels on its own, without an engineer writing out the joint angles for every step.

Related

Imitation Learning · Reinforcement Learning · Policy · Sim-to-Real Transfer · Foundation Model · Embodied AI

Sources

Robot learning - Wikipedia