The Embodiment Gap in Robot Foundation Models

AI summary

Overview

Research area: Robotics — robot foundation models (RFMs), vision-language-action (VLA) policies, cross-embodiment learning, and real-robot deployment.

Technical level: Intermediate. This is a conceptual survey that assumes familiarity with VLA policies, generalist robot policies, teleoperation interfaces, and robot data formats, but it introduces no new mathematics or empirical benchmark results of its own.

Scope (one sentence): The paper surveys what can be reused across different robot bodies and what work still remains before a shared model, representation, or dataset can actually be executed on a specific target robot.

What This Paper Is About

Robot foundation models are usually discussed through a scaling lens — more data, bigger models, broader benchmarks should yield better generalization. The authors argue that in robotics a model can generalize in the abstract while a large amount of engineering work still remains before it runs on a robot with a particular body, controller, gripper, or installation. They name this gap the embodiment gap and set out to catalog which parts of the gap scaling can reduce and which parts remain as hands-on work on the target robot.

Key Contributions

1. A definition of the embodiment gap. The authors define it as the gap that arises when models, representations, or data reused across robots must be converted into executable actions matching the body and control system of a target robot. A problem counts as part of the gap when additional work is required on the target robot after transfer across embodiments.

2. A two-axis qualitative map of existing work. The horizontal axis orders shared structure by its distance from execution (semantics and tasks, perception and affordances, object interaction, actions and skills, morphology-aware sharing). The vertical axis shows where work primarily remains on the target robot (connecting plans to skills, calibration and control-interface alignment, contact and force execution, safety and recovery). Representative methods are placed on this map.

3. An organization of the field into three overlapping research directions: sharing semantics and perception, sharing robot data and interfaces, and learning correspondence across embodiments. The map shows each direction occupying a different region and leaving different residual work.

4. A reporting framework for adaptation work. The framework is intended to record the work behind a reported success rate — what had to be done on the target robot, and what failure causes arose — because success rate alone does not reveal deployment cost or difficulty.

Main Findings

• Perceived scaling and reuse do not equal executability. The paper's central claim is that a shared model or representation does not directly become motion on a real robot; it must be connected to the target robot's body and control system. The practical value of two systems can differ even when they reach the same success rate, because they may require different amounts or kinds of work first.

• Work moves closer to physical execution as shared structure does, but not one-to-one. The authors contrast OpenVLA and UMI / ALOHA: both share actions or skills, but OpenVLA still requires its outputs to be aligned with the target robot's controller, whereas UMI / ALOHA leaves the problem of reproducing collected motion as stable contact on the target robot.

• Coding reliability differed sharply across the two axes. Three researchers independently coded 21 systems or method groups. On the shared-structure axis, all three coders agreed on 16 groups (76.2%) and at least two agreed on all 21 groups (Fleiss' κ = 0.78). On the axis for where work remains, all three agreed on 10 groups (47.6%) and at least two agreed on 20 groups (95.2%; κ = 0.33). The coders agreed on both coordinates for 7 groups (33.3%).

• The calibration-versus-contact boundary is the hard judgment call. Ten of the 11 disagreements on the vertical axis were between adjacent categories, mostly between control-interface alignment and the contact-rich execution that follows. XSkill / UniSkill was the only group for which the three coders selected three different stages. As a result, the authors added a priority rule for adjudication.

• Safety and recovery is a visible gap. No method group among the 21 was assigned to the Safety / Recovery row after adjudication. The authors state this does not mean safety and recovery are solved — rather that, within the survey's scope, safe stopping and recovery have not yet become a central research target of robot foundation models.

• Shared semantics and perception stop short of motion. Methods such as SayCan, CLIPort, PerAct, R3M, RT-Affordance, RoboPoint, VoxPoser, and Perception Stitching can indicate what to do or where to act, but they do not determine how the target robot approaches a location, maintains contact, or adapts to a change in reachability or gripper shape.

• Video and world models move closer to execution. Because they represent change over time, video-based methods can narrow the range of actions a robot considers, but they still mainly specify a desired change; producing that change on a particular robot and correcting drift during contact still has to be designed.

• Robot data sharing produces shared formats, not shared physical behavior. Datasets and infrastructures including RoboNet, BridgeData V2, Open X-Embodiment / RT-X, DROID, RLDS, LeRobot, RoboMIND, OXE-AugE, and RoboWheel make robot experience a reusable training resource, and schemas and normalization (for example SE(3) transforms, object-centered frames, equivariant representations, action tokenization) make heterogeneous data easier to train on jointly. The authors emphasize that a common action representation does not guarantee common physical execution — the same command can produce different contact or pressure on a different body.

• Teleoperation interfaces bake in an embodiment. ALOHA, DROID, UMI, OPEN TEACH, and AnyTeleop make consistent collection easier, but each assumes a particular gripper, input device, or motion pattern; replaying recorded motion on a robot with different arm length or gripper can place a goal out of reach, cause a collision, or change contact.

• Benchmarks under-report the process. LIBERO and RLBench standardize tasks and success criteria, but existing evaluations usually emphasize final success rate and often provide little detail about camera or frame readjustment, human intervention to stop the robot, or manual resetting after failure. The authors argue similar success rates therefore do not establish similar deployment conditions.

• Morphology-aware and force/tactile methods narrow but do not close the gap. Body Transformer, GET-Zero, X-VLA, morphology-aware Transformers, and RodriNet supply body structure or kinematics so action generation can depend on the embodiment, which can reduce simple motion mismatch. Knowing kinematic structure does not determine how a real gripper presses or slips; contact depends on end-effector shape and material and on the real controller's response. Tactile and force work (TactAlign, UniTacHand, Feel the Force, ForceMimic, Tactile-VLA, TaF-VLA) makes contact deviations observable when vision alone is insufficient.

• A noted disclosure: the authors state they used GPT-5.5 to proofread the manuscript and prepare a draft of Figure 1, and take full responsibility for the final manuscript.

• Not reported: this survey does not report its own benchmark experiments, success rates, or quantitative deployment measurements. The only quantitative results in the provided content are the inter-coder agreement statistics above.

Methodology in Plain English

The authors did not run robot experiments. They wrote a survey with four methodological components:

1. Definition and boundary setting. They define the embodiment gap operationally and give a practical inclusion test: does the problem arise when a shared model or representation is executed across different robot embodiments? Changes in lighting, object appearance, or task procedure on the same robot are treated as domain shift or task adaptation and kept outside the main scope, unless they are tied to deployment on another robot. Appendix A covers survey scope and literature selection; Appendix B gives contrasting inclusion and exclusion examples.

2. Two-axis qualitative mapping. They build a horizontal axis of shared-structure type and a vertical axis of where work remains, then place representative methods on the map. A system combining several kinds of shared structure is placed by the structure supporting its main transfer claim. Where data infrastructure has no dedicated category, methods are placed under action and skill when their trajectories or actions are reused as a training resource; Appendix C records these forced choices and placement rationales.

3. Independent coding and adjudication. Three researchers independently coded 21 systems or method groups using a shared initial codebook, selecting one primary category per axis plus a secondary candidate, a confidence rating, and a boundary-case flag. Boundary cases were adjudicated after rereading each method's main claims and the target-robot work identifiable from public materials. The final coordinates are stored in one machine-readable table used by both Figure 2 and Table 7. Appendix D provides a compact reporting checklist.

4. Synthesis into three directions and a reporting framework. The map's translucent regions group methods into the three overlapping research directions examined in Sections 4 through 6, and Section 7 converts the identified gaps into reporting items.

Why This Matters

Impact on research. The paper reframes how cross-embodiment and VLA results should be read. Instead of asking only "did the robot succeed," it asks "how much work was required on the target robot first, and where does that work sit." It also introduces a coding-reliability result showing that researchers can agree fairly well on what is shared (κ = 0.78) but agree much less on where residual work lies (κ = 0.33) — evidence that the field lacks shared vocabulary for the deployment side of robot learning. The absence of any Safety / Recovery placement among the 21 groups is presented as an exposed research gap.

Real-world applications (as grounded in the paper):

• Industrial robotics deployment. The authors note that system integrators in industrial robotics often carry out much of this less visible work through experience and tacit knowledge — the same kind of work that appears when an RFM is deployed on a new body.
• Bringing a trained policy to a new robot arm or body. The paper's running example is using the same policy on a different arm and having to adjust the control system or action interface, and the effect of a changed number of fingers on grasp and contact pattern.
• Teleoperation and data collection at scale. Understanding how interfaces such as ALOHA, DROID, UMI, OPEN TEACH, and AnyTeleop constrain later transfer matters for teams deciding how to collect physical robot data.
• Benchmark and evaluation design. The proposal that benchmarks record the process (camera or frame readjustment, human stops, manual resets) as well as final performance applies directly to how LIBERO- and RLBench-style evaluations, and real-robot evaluations, are reported.

Industry relevance. The paper speaks to the practical economics of deploying foundation-model policies: two systems with the same success rate can require very different amounts of integration, calibration, and recovery engineering, which determines deployment cost and whether a robot can run continuously in the real world when the stopping and recovery process is unclear.

Future Directions

• Bring adaptation work into the model or system design. The authors state that future systems need to incorporate more of the target-robot adjustment into the model or system design rather than leaving it as human effort.

• Report the human work behind real-robot deployment. Papers should state where human work was required during real-robot deployment, and benchmarks should record the process that made performance measurement possible, not just the final number.

• Connect temporal video knowledge to closed-loop control. The authors identify this as an important open problem for turning video and world-model predictions into reachable motion with correctly aligned contact timing and speed.

• Establish what data and representations to scale for cross-embodiment correspondence. The field has not yet determined which data should be collected or which representations should be scaled for RFMs that generate actions across several embodiments while accounting internally for their differences.

• Make safety, stopping, and recovery a central research target. The absence of any Safety / Recovery placement suggests one possible architecture: a generalist policy generating motion while a higher-level agent decides when to stop or recover.

Target Audience

This paper is most useful to robotics and robot-learning researchers working on VLA policies, generalist robot policies, and cross-embodiment transfer; to engineers and system integrators who must actually deploy such models on a specific robot; and to benchmark and evaluation designers who want reported success rates to be accompanied by the adaptation work behind them. It is also valuable to research managers and industry teams weighing the practical cost of moving a foundation-model policy onto a new body. Readers without prior exposure to robot learning terminology will find it accessible conceptually but will benefit from knowing what VLA policies and teleoperation interfaces are.

Authors’ abstract

Robot foundation models (RFMs), including vision-language-action (VLA) policies, are often discussed through a scaling view: more data, larger models, and broader benchmarks should improve generalization. In robotics, however, a model can generalize while work still remains before it can run on a robot with a particular body. The work required differs across methods and target robots, and those differences affect practical deployment. We call the gap between reusable models, representations, or data and their use in execution on the target robot the embodiment gap. This survey examines what can be reused across robot embodiments and what must still be implemented on a new robot. We place existing methods on a two-axis map that shows the type of shared structure and the stage at which adaptation is needed for execution on the target robot. We then examine recent work through three overlapping research directions: sharing semantics and perception, sharing robot data and interfaces, and learning correspondence across embodiments. We also propose a reporting framework for adaptation work that success rate alone does not reveal. The framework identifies the work that should be checked when comparing cross-embodiment learning and highlights work that remains on a new robot and questions for future study.