General-purpose robots must work out what a new task demands and turn that understanding into physical action. A newly posted survey on Hugging Face, In-Context Learning for Robots: Methods and Applications, examines how demonstrations and interaction can steer a robot's existing competence while its neural parameters stay fixed during deployment.
The review is organized around the interfaces that connect contextual evidence to execution, and it distinguishes four families: context-conditioned policies, geometric demonstration transfer, world-model-based control, and skill- and agent-based execution. Comparing these interfaces, the authors argue, clarifies the transfer assumptions each one makes and the distinct roles played by training, correspondence, and memory in making context useful.
Across both manipulation and navigation, the survey tracks how these mechanisms preserve taught requirements as objects, environments, and execution conditions change. That framing separates two questions: whether the robot correctly inferred the requirement supplied by the teaching, and whether its executor can realize that requirement once the setting shifts.
The authors link method design to evaluation practices intended to distinguish responsiveness to teaching, physical transfer, and gains from retained experience. They also separate fixed-weight contextual adaptation from related fast-weight approaches, and note that the evaluation chapter may serve as a practical starting point for readers designing experiments.
The paper runs to 100 pages with 412 references, and a companion repository organizes a growing reading list by taxonomy and first public release date. The survey's stated agenda connects compositional task acquisition and faithful transfer to what it calls physical recursive self-improvement, in which experience improves a robot's ability to learn subsequent tasks.