Behavior Cloning
Also known as BC
Overview
Behavior cloning learns a policy by matching expert actions to observed states or images, offering a direct way to turn demonstrations into control without designing a reward function. Common approaches include supervised regression, sequence and diffusion policies, visual or language-conditioned models, and methods that retrieve, reweight, or augment demonstrations. Researchers evaluate these methods in settings such as simulated driving, robotic manipulation, navigation, and long-horizon tasks, with growing attention to recovery from errors and shifts in the environment. A central challenge is that good action-prediction scores do not guarantee successful closed-loop behavior: errors can compound, demonstrations may be noisy or incomplete, and policies must handle multimodal actions and unfamiliar states safely.
Latest papers
54 papers9 papers in the last four weeks, down 25% on the four weeks before. 0.1% of all new papers.