LfD

LfD, short for Learning from Demonstrations, is a family of techniques in machine learning and artificial intelligence that enables an agent to acquire tasks by observing demonstrations from an expert. It is closely related to imitation learning and apprenticeship learning, and it is widely used in fields such as robotics, autonomous systems, and AI research.

In LfD, demonstrations provide traces of states and actions (or sequences of states) collected from humans, simulators, or scripted controllers. The learning objective is to reproduce the demonstrated behavior. Behavior cloning trains a policy that maps states to actions directly from labeled data. Inverse reinforcement learning seeks to infer the reward structure that produced the expert's behavior, after which a policy is derived by optimizing that reward. Apprenticeship learning combines elements of IRL with performance constraints to avoid mimicking suboptimal strategies. Some approaches also employ interactive or query-based methods to improve data efficiency.

Methods and settings vary, with some LfD techniques using end-to-end deep learning and others relying on structured models or probabilistic frameworks. Common challenges include data efficiency, generalization to states not seen in the demonstrations, distributional shift between training and deployment, and ensuring safety during learning and execution.

Applications of LfD span robotic manipulation, autonomous driving, healthcare robotics, and simulation-to-real-world transfer. Limitations often relate to the quality and coverage of the demonstrations, the complexity of the task, and the difficulty of accurately modeling environment dynamics.

See also: Imitation learning, Behavior cloning, Inverse reinforcement learning, Apprenticeship learning.