"The robot learns from demonstration" is the phrase of the moment, and like most phrases of the moment it hides its mechanism. Two grants strip it down. Naver's US12632695B2, "Robotic demonstration retrieval systems and methods" (issued May 19, 2026), and Figure's US12638859B2, "Bipedal action model" (issued May 26, 2026), together describe the full loop — and its limits.
What the filings actually say: a robot is shown example behaviors (a human teleoperating it, or guiding its limbs). Those demonstrations are stored. Naver's patent is specifically about retrieval — given a new situation, find the most relevant stored demonstrations to draw on. Figure's action model is about the other half: learn a policy that, conditioned on perception and goal, predicts actions that imitate what the demonstrations did. Show, store, retrieve, imitate. That's the mechanism the buzzphrase compresses.
Strip the marketing and you see the dependency the phrase hides: the robot is only as good as its demonstrations. Retrieval (Naver) matters precisely because coverage is finite — when the current situation resembles a demonstration, the robot does well; when it doesn't, retrieval returns a poor match and the imitation policy extrapolates into territory it was never shown. The capability is genuine, but it is bounded by the demonstration set, not by anything like general competence.
The boring incumbent here is hand-engineered control — slow to build, but predictable everywhere it's defined. Learning from demonstration trades that predictability for fast capability inside the demonstrated distribution and unpredictable behavior outside it. That's a real and often worthwhile trade. It is not the same as a robot that "understands" the task, and the two patents make the distinction concrete: there is retrieval and imitation, not comprehension.
What neither grant discloses is the out-of-distribution failure rate — how often the robot meets a situation unlike anything demonstrated, and what it does then. That number is the whole ballgame for deployment, and it lives in evaluation data no method claim contains. A patent on retrieval tells you the system tries to find a relevant precedent; it does not tell you what happens when there isn't one.
No claim about a capability ships here without a primary-source anchor, so here is the anchor: "learning from demonstration" is, per these filings, show-store-retrieve-imitate. Read that way, the buzzphrase becomes falsifiable — and the right question to ask any robotics company stops being "does it learn?" and becomes "what did you demonstrate, and what happens off the script?"