Robotic in-context learning (ICL) is an emerging paradigm in which a robot infers and executes tasks from visual demonstrations. But as researchers behind a new paper point out, the problem itself has remained under-defined: a single visual demonstration can convey action trajectories, object semantics, manipulation affordances, spatial relations, and task goals all at once, leaving it unclear what information the robot is actually expected to follow.
To resolve this ambiguity, the authors first provide a clear problem definition of robot ICL that explicitly states its learning target. Building on that definition, they develop SimpleICL, a minimalist and reproducible framework that combines a visual prompt encoder with a low-cost data collection protocol.
According to the paper, the framework achieves strong performance in both simulation and real-world environments without massive pre-training or specialized data infrastructure. Experiments across simulation and eight real-world tasks reportedly show strong zero-shot generalization and robustness, and reveal key properties of robot ICL, including action, semantic, composition, and affordance discrimination.
The authors state that they will fully open-source their data and training pipeline to support systematic and reproducible research on robot ICL. In discussion, they confirmed that the dataset and training code will be released soon. A separate survey on in-context learning for robots has also noted plans to discuss SimpleICL in a future version.