Imitation learning (also called learning from demonstration) is a training paradigm where a robot acquires skills by watching how a task should be done rather than discovering behaviors through trial and error. A human operator demonstrates the desired behavior — either through teleoperation, kinesthetic teaching (physically moving the robot's arms), or even video recordings — and the robot learns a policy that maps sensory observations to actions that replicate the demonstrated behavior.
This approach has gained enormous momentum in humanoid robotics. Companies like Physical Intelligence, Toyota Research Institute, and Google DeepMind have shown that transformer-based imitation learning policies, trained on thousands of human demonstrations, can produce robots capable of folding laundry, loading dishwashers, and manipulating diverse objects. The appeal is clear: rather than engineering behaviors from scratch, you show the robot what to do and let it generalize.
The main challenges are data collection (gathering enough high-quality demonstrations is labor-intensive and expensive), distribution shift (the robot encounters situations the demonstrator never showed), and compounding errors over long task horizons. Active research focuses on scaling demonstration datasets, combining imitation learning with reinforcement learning for self-improvement, and building foundation models that can generalize from a modest number of demonstrations to a wide range of tasks. For deeper coverage, see HumanoidIntel.