- The paper defines the embodiment gap as the distance between reusable models, representations, and data and executable motion on a target robot, including controller alignment, contact execution, and safety.
- The survey maps 21 method groups across shared structure and residual deployment work, finding stronger agreement on what transfers than on the engineering work that remains.
- The paper proposes report cards and Embodiment Adaptation Curves to document demonstrations, calibration, interventions, resets, and failures that conventional success rates often conceal.
This survey, "The Embodiment Gap in Robot Foundation Models" (2608.18433), examines a problem that success-rate reporting in robot learning tends to obscure: a model, representation, or dataset that generalizes across robots still cannot execute on a particular body without substantial additional work. The authors—researchers at AIST, Japan—define this embodiment gap as the distance between reusable structure (models, representations, data) and executable motion on a target robot's body and control system. The paper contributes an operational definition of the gap, a two-axis qualitative map of existing methods coded independently by three researchers, a synthesis of three research directions, and a reporting framework designed to expose adaptation work that final performance metrics hide.
Motivation and positioning
The survey responds to the scaling narrative dominant in vision-language-action (VLA) policy research—RT-X, Octo, OpenVLA, RoboCat, and the π0​ family—by arguing that scaling data and model size does not eliminate the work required to connect shared structure to a specific robot's kinematics, controller, end effector, and contact behavior. The authors align themselves with Goldberg's argument that robotics lacks Internet-scale action-paired observations and that engineering will remain essential regardless of data volume, and with the related Science Robotics debate over whether "data will solve robotics." Their contribution is not to reject scaling but to ask precisely which portions of the deployment workload can be reduced by scale and which persist as target-robot engineering.
The scope is deliberately bounded: manipulation-focused RFMs and VLA policies, plus surrounding infrastructure (teleoperation interfaces, datasets, benchmarks, evaluation systems). Domain shift on a single robot, task adaptation, and sim-to-real transfer are excluded unless they co-occur with a change of embodiment; system integration generally is excluded except for the portion needed to make a transferred policy executable on new hardware.
A two-axis map of what is shared and where work remains
The paper's central analytic device orders shared structure along a horizontal axis by distance from execution—semantics and tasks, perception and affordances, object interaction, actions and skills, and morphology-aware sharing—and places residual deployment work along a vertical axis: skill/API connection, calibration and control-interface alignment, contact and force execution, and safety/recovery. Twenty-one method groups were placed on this map through independent coding by three researchers using a shared codebook, with adjudication of boundary cases.
The coding statistics are informative about the taxonomy itself. On the shared-structure axis, all three coders agreed on 16 of 21 groups (76.2%; Fleiss' κ=0.78). On the remaining-work axis, agreement was markedly weaker: unanimous on only 10 groups (47.6%), κ=0.33, though at least two coders agreed on 95.2% of cases. Ten of eleven vertical-axis disagreements fell between adjacent categories, concentrated on the boundary between control-interface alignment and subsequent contact-rich execution—a boundary the authors acknowledge requires judgment and which motivated a priority rule added during adjudication. The authors state plainly that this check "does not establish that the categorization is fully objective"; it shows only that the horizontal axis is comparatively reproducible while vertical placement involves interpretive decisions.
Two findings from the map carry substantive weight. First, no method group was placed in the Safety/Recovery row. The authors explicitly caution this does not mean safety and recovery are solved; rather, safe stopping and post-failure recovery have not become central research targets of RFM development, despite being necessary for sustained real-world operation. Second, placement reveals a non-one-to-one relationship between shared structure and residual work: OpenVLA and UMI/ALOHA both share actions or skills, yet OpenVLA leaves controller-alignment work while UMI/ALOHA leaves the problem of reproducing collected trajectories as stable contact. What remains depends on how far each design extends into execution, not merely on the abstraction level of what it shares.
Three research directions
Sharing semantics and perception leverages web-scale language, image, and video data. Methods such as SayCan decompose tasks and select skills but presuppose that executable skills already exist on the robot; affordance methods (CLIPort, PerAct, RT-Affordance, RoboPoint, VoxPoser) identify where to act but not how to approach or maintain contact. Video prediction and world-model methods move closer to execution by representing temporal scene change, yet they specify desired changes rather than the motions realizing them under closed-loop control. Large VLA backbones combine these cues with multi-robot data, but their action representations inherit the control conventions of training robots—the same action token can produce different motion and contact when the target controller interprets magnitude or rate differently, and fine-tuning studies confirm that target-robot data are typically needed to restore alignment.
Sharing robot data and interfaces standardizes physical experience through datasets (Open X-Embodiment, DROID, BridgeData V2), schemas (RLDS, LeRobot), teleoperation platforms (ALOHA, UMI, OPEN TEACH), object-centered and equivariant action representations, diffusion-based trajectory generation, frameworks (PyRobot, LeRobot), and benchmarks (LIBERO, RLBench, CALVIN). The recurring limitation the authors identify is that common formats do not guarantee common physics: an identical end-effector displacement produces different contact pressure on a different arm or gripper, and a trajectory natural in training data can be unstable when camera placement or controller response changes at deployment. Benchmarks compound the problem by emphasizing final success rates with little record of calibration frequency, manual intervention, or resets preceding evaluation.
Learning correspondence across embodiments models relations between bodies directly. High-level correspondence shares task progress or skill meaning (XSkill, UniSkill); intermediate correspondence shares object or scene change via flow, point tracking, and latent actions or action motifs (Track2Act, MOTIF); morphology-aware methods condition policies on body graphs and kinematics (Body Transformer, GET-Zero, X-VLA, RodriNet); and tactile/force methods (TactAlign, Tactile-VLA, ForceMimic) make contact deviations observable. The authors' assessment is consistent across levels: knowing kinematic structure does not determine how a gripper presses against an object or how much it slips; tactile signals' magnitudes depend on sensor placement and must themselves be aligned across embodiments. Deformable objects, control delays, and varying sensor configurations further resist collection at scale.
Reporting adaptation work behind success rates
The most actionable contribution is a reporting framework with three components: a minimum report card tracing the path from source system to final evaluation, Embodiment Adaptation Curves (EACs) relating performance to measured adaptation effort, and a four-way failure-cause taxonomy mapped to the research directions.
Applying the report card to representative cross-embodiment claims yields stark findings. Among nine method groups examined, counts of safety interventions and recovery events could not be identified from papers or supplementary materials for seven; setup and calibration conditions were insufficiently documented for three. For OpenVLA, worked out in detail in an appendix, target-robot demonstrations (10–150 per Franka task), fine-tuning scope (full FT or LoRA), and evaluation rollouts are identifiable, while adaptation trials, resets, robot time, intervention counts, calibration metadata, and systematic failure causes are not. The implication drawn is direct: readers cannot determine whether a reported success rate reflects autonomous capability or researcher effort, so identical success rates may mask very different deployment costs.
EACs address this by plotting performance against a comparable measure of adaptation effort—demonstrations, real-robot trials, or human interventions. MOTIF serves as a rare full-curve example because it reports few-shot transfer at multiple demonstration counts; Track2Act admits only an endpoint comparison (open-loop versus a residual policy trained on 400 Spot teleoperation trajectories), which the authors include while conceding that the information needed for full curves is frequently unavailable in published results.
Limitations and open questions
Several limitations are acknowledged within the paper itself. The two-axis map is qualitative and explicitly not predictive—it estimates neither the amount of work nor ranks methods—and its vertical axis showed weak inter-coder reliability. The literature set is a scoping survey weighted toward manipulation; extension to locomotion and navigation, where terrain, dynamics, and long-horizon autonomy dominate, is proposed but unexecuted. Recent preprints and OpenReview submissions support directional claims, with the authors distinguishing accepted from unreviewed work and noting reliance on published results for central claims. The report-card application depends entirely on public documentation, so N/R entries conflate genuinely absent reporting with reporting in inaccessible venues.
Open questions left explicit include: which data and representations should be scaled for embodiment-correspondence models capable of generating actions across embodiments internally; how video-derived temporal knowledge can be connected to closed-loop contact control; whether contact and force representations can be made shareable across differing sensor configurations; and how failure detection, safe stopping, and recovery should be architected—whether within the policy, the controller, or a supervisory agent.
Conclusion
The survey reframes progress in robot foundation models around a question that aggregate success rates suppress: after a model transfers, what had to be built, calibrated, collected, intervened upon, and recovered on the target robot before that number existed? Its definition of the embodiment gap, its coder-audited map showing residual work migrating toward physical execution as shared structure approaches it, and its report card and EAC proposal together supply a vocabulary and measurement practice for making deployment cost visible. The empty Safety/Recovery row and the pervasive N/R entries in its own case studies indicate that the field's current reporting practices leave the most consequential deployment work undocumented—an omission the authors argue should be corrected as a matter of standard practice rather than left to inference.
