A 100-page, 412-reference survey splits robot in-context learning into four interfaces and argues for evaluating 'did it infer the teaching' separately from 'can it execute after objects change'

The survey proposes interfaces as the organizing axis and divides robot ICL into four families: context-conditioned policies, geometric demonstration transfer, world-model-based control, and skill- and agent-based execution. Earlier surveys organize learning from demonstration by teaching interfaces and learned policies, rewards, or plans, or organize manipulation ICL by context content, inference targets, adaptation mechanisms, and transfer; this one instead asks through which intermediate the contextual evidence reaches execution. This is a survey-level taxonomic contribution: the abstract and introduction state the four families and argue that comparing these interfaces clarifies their transfer assumptions and the roles of training, correspondence, and memory in making context useful, with manipulation and navigation as the stated scope.

The survey separates two questions: whether the robot inferred the requirement supplied by the teaching, and whether its executor can realize that requirement after the objects or environment change. This distinction separates broader motor competence from broader ability to learn through teaching, noting that broader pretrained motor competence, more informative teaching, and selective experience reuse address different limits of adaptation. The abstract and introduction state the distinction and give the example that placing an object at the correct destination can still violate a required handle grasp, so both the intended effect and the prescribed order or contact must be traced through execution; cross-object transfer is described as testing whether teaching remains useful after replacing either or both interacting objects.

The survey links method design to evaluation practices, arguing that evaluation should distinguish responsiveness to teaching, physical transfer, and benefits from retained experience. Relative to complementary reviews emphasizing data and evaluation or cross-embodiment adaptation, this work treats context dependence, transfer, and retained-experience benefits as objects to be identified separately under evaluation controls, and builds an agenda of compositional task acquisition and improved teachability on that basis. The abstract states that the analysis links method design to evaluation practices; the introduction lists as a third contribution a synthesis of reported comparisons and evaluation controls; the community note says readers designing experiments may find the evaluation chapter a useful starting point.

The survey offers a longer-term agenda connecting compositional task acquisition and faithful transfer with physical recursive self-improvement, in which experience improves the ability to learn subsequent tasks. The agenda treats retained context, executable programs, and neural updates as different channels through which physical self-improvement can operate, and adds collective knowledge evolution as an objective extending the learning cycle to a group of robots. The abstract and introduction frame this through the S5 and S6 learning horizons and note that realization can combine several earlier capabilities; the text presents these as research objectives rather than achieved results.

The survey addresses researchers and practitioners who need general-purpose robots to infer new task requirements from demonstrations, corrections, and interaction; its scope is manipulation and navigation, focused on settings where neural parameters are held fixed during deployment and context directs existing competence. Its taxonomy and evaluation suggestions are meant for designing experiments and comparing methods, including transfer tests under object substitution, unfamiliar environments, and changing execution conditions, and evaluation controls that separate context dependence, physical transfer, and retained-experience benefits. Physical recursive self-improvement and collective knowledge evolution are framed as research objectives whose realization can combine several earlier capabilities.

As a survey, its conclusions take the form of literature synthesis and taxonomy rather than results from a single experiment, so the relative merits of the four families depend on the settings of the cited studies. Physical recursive self-improvement and collective knowledge evolution are listed as research objectives whose feasibility remains to be tested by later work. In addition, the readable text here consists of the abstract, introduction, and portions of the body; figures and tables are not fully presented, so method-level details of the four families and the concrete evaluation controls still need to be confirmed in the corresponding sections of the original.