The Embodiment Gap in Robot Foundation Models
Yukiyasu Domae, Keisuke Shirai, Hanbit Oh, Ryoichi Nakajo, Tomohiro Motoda, Koshi Makihara, Masaki Murooka, Takuma Yagi, Yoshiaki Bando, Ryo Hanai
cs.RO, cs.LG
2026-08-19
AIST defines the embodiment gap as leftover work to run a shared robot model on a new body, maps 21 method groups, and asks papers to report adaptation effort behind success rates.
Robot foundation models, especially vision-language-action (VLA) policies, are usually cast in the same scaling story used for language and vision: more data, bigger models, broader benchmarks, better generalization. Robotics adds a second bill. A model can look general on paper and still need calibration, control-interface alignment, contact handling, and failure recovery before it runs on one particular body.
Industrial robot deployment has always hidden that bill inside system-integrator craft. The same bill shows up when OpenVLA, Octo, or π0 is moved onto a new arm, often under new cameras and controllers. Goldberg's point is blunt: language and vision can train on internet-scale observations, while robotics cannot collect observation-action pairs at a similar scale. A Science Robotics debate treats the split between data, models, and engineering as still open. This AIST survey asks which parts of the leftover work shrink when data and models scale, and which parts stay as engineering on the target robot.
The embodiment gap is the leftover work that appears when a model, representation, or dataset reused across robots has to become executable motion that matches a target robot's body and controller. Lighting changes, new tasks, and sim-to-real on the same body sit outside the definition. A different arm, gripper, or action interface sits inside it.
Existing methods are placed on a two-axis qualitative map. The horizontal axis orders what is shared, from far from execution to close: semantics and tasks, perception and affordances, object interaction, actions and skills, morphology-aware sharing. The vertical axis orders where work remains on the target robot: skill or API connection, calibration and control-interface alignment, contact and force execution, safety and recovery.
Three researchers independently coded 21 method groups. On the shared-structure axis, all three agreed on 16 of 21 groups (76.2%, Fleiss' κ = 0.78). On remaining work, full agreement dropped to 10 of 21 (47.6%, κ = 0.33). Ten of the 11 vertical disagreements were between adjacent categories, mostly calibration versus contact. XSkill / UniSkill was the only group that received three different stage labels.
Three overlapping research directions occupy different regions of the map.
The survey also proposes an 11-item report card and the Embodiment Adaptation Curve (EAC). The curve plots performance against a comparable unit of leftover work, such as target-robot demonstrations, real-robot trials, or human interventions.
This is a survey, not a new policy bake-off. The hard numbers come from coding agreement and from what published papers omit.
| Item | Number |
| Method groups coded | 21 |
| Full agreement on shared structure | 16/21 (76.2%, κ=0.78) |
| Full agreement on remaining work | 10/21 (47.6%, κ=0.33) |
| Full agreement on both axes | 7/21 (33.3%) |
| Groups placed on Safety / Recovery | 0/21 |
| Report-card sample with unidentifiable safety/intervention counts | 7/9 |
| Sample with unidentifiable calibration detail | 3/9 |
On individual systems, OpenVLA fine-tunes on Franka with 10-150 demonstrations per task across seven tasks, using full fine-tuning or rank-32 LoRA. RoboCat reports 100-1000 examples. MOTIF studies 1-50 shot cross-embodiment transfer. Track2Act trains a residual policy on 400 Spot teleoperation trajectories. A full EAC can be sketched from MOTIF's multi-shot table. Track2Act supplies only an open-loop versus residual-policy endpoint. Most papers in the survey do not even give that.
A κ of 0.33 on the remaining-work axis means the line between interface alignment and contact execution is itself unstable. An empty Safety / Recovery row does not mean the problem is solved. Within this literature, safe stop and resume has not become a primary research target of robot foundation models.
For anyone who wants to run a VLA on a new body, the survey splits a success rate into inspectable deployment cost. Two systems can post similar numbers while one swaps an action head and a few dozen demos, and the other runs full fine-tuning plus frequent human e-stops. The report card asks for source robot, target robot, reused structure, modified components, target-robot data, model-update scope, calibration, real-robot operation, safety interventions, evaluation rollouts, and failure causes. The tacit work of industrial integration finally has a list that can go in a paper.
This is a framing paper, not a new algorithm. It will not pick next week's VLA. It tells you which questions to ask when two cross-embodiment numbers look alike. If the field keeps reporting success rate alone, the embodiment gap stays buried in lab procedure.
The two-axis map is qualitative by design: it does not estimate hours of work and does not rank methods. Coordinates for the 21 groups come from independent coding plus later adjudication; vertical agreement is weak, and the priority rule that separates calibration from contact was added after the first pass. Coverage leans toward manipulation and VLA. Locomotion, flight, and soft robots are largely out of scope. An N/R in Table 2 means the item could not be recovered from public materials, not that the original authors skipped the work. Almost no paper currently supports a full EAC, so the framework still lacks evidence of community uptake. The authors used GPT-5.5 to proofread the manuscript and draft Figure 1, so the figure should be read as a human-owned schematic rather than a model-generated claim.
Four future priorities are stated clearly: record how an action became executable, not only the successful trajectory; design RFMs as systems that help adapt to the target robot; push contact-rich execution toward structures that can be shared across bodies; and count the work required to obtain a success rate as part of evaluation. Those are an agenda, not a validated method.