Robot foundation models hit 'embodiment gap' — new survey maps adaptation work

Robot foundation models (RFMs) — including vision-language-action (VLA) policies — are often marketed on a scaling promise: feed them more data and larger parameters, and generalization will follow. But a new survey from a team led by Yukiyasu Domae (with 9 co-authors) argues that robotics has a blind spot that scaling alone won't fix. The paper, published in Transactions on Machine Learning Research (TMLR) and posted on arXiv (2608.18433), formalizes the 'embodiment gap': the hidden engineering required to take a reusable model, representation, or dataset and make it actually run on a particular physical robot.

The team analyzes the problem through a two-axis map: what type of structure is shared across embodiments, and at what stage adaptation is needed for execution on the target robot. They categorize recent work into three research directions — sharing semantics and perception, sharing robot data and interfaces, and learning correspondence across embodiments. Crucially, they propose a reporting framework that goes beyond success rate metrics, forcing comparisons to account for the remaining per-robot work. This shifts the conversation from 'does it generalize in a benchmark?' to 'how much integration effort does a new robot body require?' — a question that determines whether foundation models are practical for real-world deployment.