Skip to main content
Back

AIST Survey Exposes the Hidden Work Behind Robot Foundation Models

A new TMLR survey argues that success rates hide the real cost of moving robot foundation models to new hardware, and proposes a reporting framework to fix it.

Paper, Full Access
  • New TMLR survey introduces the embodiment gap: reusable models still need work to run on a specific robot.
  • Two-axis map places 21 systems by what they share and what adaptation remains on target hardware.
  • Coder agreement was high on shared structure (kappa=0.78) but low on remaining work (kappa=0.33).
  • Zero of 21 method groups tackle safety and recovery as a central research target.
  • Proposes a report card and Embodiment Adaptation Curve so success rates cannot hide hidden work.
  • Argues future RFMs must integrate calibration, contact correction, and recovery into the model itself.

Robot foundation models are having their scaling moment. OpenVLA, Octo, RT-X, and the π0 family all promise generalist policies trained on massive multi-robot datasets. A new survey from researchers at Japan's AIST, published in Transactions on Machine Learning Research, makes an uncomfortable point: a model that generalizes on paper still may not run on your robot without substantial hidden work.

The authors give this hidden work a name. They call the gap between reusable models, representations, or data and their use in execution on a particular robot the embodiment gap. The survey maps where this gap appears, why success-rate numbers obscure it, and how papers should report the adaptation work behind those numbers.

Why generalization is not deployment

The scaling recipe from language and vision does not transfer cleanly to robots. Goldberg has argued that language and vision can draw on Internet-scale training data, whereas robotics cannot obtain observations paired with action commands at a similar scale, and useful robot data will only accumulate once learning and engineering combine well enough to make robots reliable.

The survey's operational definition is practical. The embodiment gap arises when models, representations, or data reused across robots must be converted into executable actions that match the body and control system of a target robot. Even seemingly innocuous mismatches cause problems. When two systems use Cartesian end-effector commands, the same displacement does not necessarily produce the same physical behavior. A different gripper or mobile base changes how the robot makes contact with an object and how it recovers after an error.