Sequential robot manipulation often requires executing novel combinations of familiar instruction components, but collecting demonstrations for every possible tuple is combinatorially expensive. The authors of this paper propose a diagnostic framework that decomposes the generalization gap into three sources: marginal instruction shift, instruction-compositional shift, and context-action shift. This decomposition lets them identify when sparse training coverage is sufficient and what structure the training set must preserve for reliable action prediction.

Their experiments show that exhaustive tuple enumeration is unnecessary: a structured subset can recover strong out-of-distribution performance when it covers action-relevant dependencies. They also find that sparse training often fails due to instruction steering rather than missing low-level skills; finetuning with only one demonstration per task improves out-of-distribution success from 0.4% to 54.7%. For semantically dependent tasks, effective coverage must capture relational structure rather than only factor diversity.

The practical takeaway is that efficient robot data collection should prioritize dependency coverage in instruction space over exhaustive task expansion. The paper includes additional results, code, and dataset on the project website.