The Embodiment Gap in Robot Foundation Models

Yukiyasu Domae; Keisuke Shirai; Hanbit Oh; Ryoichi Nakajo; Tomohiro Motoda; Koshi Makihara; Masaki Murooka; Takuma Yagi

cs.ROAug 19, 2026

The Embodiment Gap in Robot Foundation Models

Authors: Yukiyasu Domae, Keisuke Shirai, Hanbit Oh, Ryoichi Nakajo, Tomohiro Motoda, Koshi Makihara, Masaki Murooka, Takuma Yagi, +2 more

Organizations: National Institute of Advanced Industrial Science and Technology (AIST), Japan

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.

Figures & tables

Figure 1: Conceptual overview of the embodiment gap. The left side shows structures reused across robots: semantics and tasks, perception and affordances, object interaction, actions and skills, and morphology-aware sharing. The right side shows work that often remains when these structures are executed on a target robot: skill or API connection, calibration and control-interface alignment, contact and force execution, and safety and recovery. The embodiment gap arises in moving from shared structure to execution on the target robot.

Figure 2: Two-axis qualitative map of the embodiment gap. The horizontal axis shows what is primarily shared across robots. The vertical axis shows where work primarily remains when that shared structure is executed on the target robot. Each point is the final placement after independent coding by three researchers and adjudication of boundary cases. The map does not estimate the amount of work or rank methods. Within-row shifts show boundary cases and avoid label overlap; the categories in Table 7 are unchanged. No method group was assigned to the Safety / Recovery row after adjudication. The translucent regions show the three research directions discussed in Sections 4–6. The placements in Figure 2 and Table 7 come from the same coordinate table.

Figure 3: Three research directions viewed through the embodiment gap. Sharing semantics and perception broadens the meanings and visual cues available to a robot. Sharing robot data and interfaces makes robot experience easier to reuse and its conditions easier to trace. Learning correspondence across embodiments models relations among different bodies through object motion, abstract actions, morphology, and contact cues. The directions overlap and are not mutually exclusive categories.

Item	What to report	Why it matters
Source robot(s) / source setting	Robot(s), sensors, control system, pretraining setting	Establishes the source conditions
Target robot	Body, sensors, task environment	Establishes the execution conditions
Shared structure	Plans, representations, skills	Shows what is reused
Modified components	Head, adapter, decoder, control system	Shows what is changed for the target robot
Target-robot data	Demonstrations, adaptation trials, recovery data	Shows experience used for adaptation
Model updates	Frozen components, adapter/LoRA, full fine-tuning	Shows the scope and cost of updating

Table 1: Minimum report card for cross-embodiment results. The card helps readers trace work required on the target robot that success rate alone does not reveal.

System	Shared structure and connection to target robot	Target data / model update	Setup / operation	Safety / intervention / recovery	Visible work on target robot
RT-X / OXE	Multi-robot data $\rightarrow$ robot-specific actions	Large mixed dataset; shared policy	Data and control conventions	Resets and safety mostly N/R	Differences across robots and adaptation costs remain visible
Octo	Generalist policy $\rightarrow$ target interface	Target data and adaptation	Controller alignment	Intervention and recovery N/R	Connection to the target robot remains
OpenVLA	7B VLA + action tokens $\rightarrow$ FT/LoRA	10–150 demonstrations; full FT or LoRA	Control rate partly reported	Qualitative recovery; counts N/R	Decoder and control-rate work
RoboCat / $\pi_{0}$	Generalist backbone $\rightarrow$ robot/task adaptation	RoboCat: 100–1000 examples; $\pi_{0}$ : FT data	Setup N/R	Safety and resets N/R	Zero-shot and fine-tuned regimes are mixed
Track2Act	Point tracks and object motion $\rightarrow$ residual policy	400 Spot teleoperation trajectories; residual-policy BC	Depth and transform fitting	Failure modes reported; counts N/R	A target-robot residual policy is still required
MOTIF	Action motifs $\rightarrow$ few-shot transfer	1–50 shots; motif-conditioned policy	Setup largely N/R	Operation and safety N/R	Operating cost remains a reporting question

Table 2: Examples of applying the report card to representative cross-embodiment claims ( Open X-Embodiment Collaboration et al., 2023 ; Octo Model Team et al., 2024 ; Kim et al., 2025c ; Kim et al., 2025b ; Zhang et al., 2025b ; Bousmalis et al., 2024 ; Black et al., 2024 ; Physical Intelligence et al., 2025 ; Bharadhwaj et al., 2024 ; Zhi et al., 2026 ; Sferrazza et al., 2024 ; Patel and Song, 2025 ; Zheng et al., 2026a ; Zhang et al., 2026 ; Wi et al., 2026 ; Huang et al., 2025 ) . N/R denotes information that was not identifiable from the main paper or publicly available supplementary material. Rows that combine related systems summarize reporting at the group level and do not imply identical placements in Figure 2.

Figure 4: Embodiment Adaptation Curves (EACs). An EAC shows how performance changes with a reported amount of work on the target robot. The horizontal axis can use demonstrations, real-robot trials, human interventions, or another comparable measure of adaptation effort. Panel (a) gives a conceptual example, while panel (b) reproduces the few-shot transfer results reported for MOTIF ( Zhi et al., 2026 ) in arXiv:2602.13764v1, Table 3. Panel (c) uses the values reported for Track2Act ( Bharadhwaj et al., 2024 ) in arXiv:2405.01527v2 / ECCV 2024, Table 2, and compares open-loop execution with a residual policy trained on 400 target-robot teleoperation trajectories. It is an endpoint illustration related to the EAC idea rather than a quantitative adaptation curve. MG/G/CG/TG denote Mild, Standard, Compositional, and Type Generalization in Track2Act.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Boundary case	Why the case is near the boundary	Treatment in this survey
Body Transformer	This policy uses body structure and sits outside the usual VLA definition.	Included as morphology-aware sharing of actions and skills.
GET-Zero	It generates policies conditioned on a body graph.	Included as a morphology-aware method.
RodriNet	This kinematic mechanism lies outside a narrow definition of an RFM.	Included as a correspondence mechanism close to execution.
Tactile and force methods	These methods may lie outside VLA and directly address contact information.	Included because they make contact-rich execution visible.
Common APIs and control interfaces	These infrastructure components determine whether shared structure can be executed.	Included as connections needed to run shared structure on a target robot.
Object flow and point tracking	They can be read as visual cues or as correspondence across embodiments.	Included as methods that share changes in objects or scenes.

Appendix

Table 3: Boundary cases and inclusion decisions.

Concept	Example outside the main scope	Example treated as part of the embodiment gap
Domain shift	Only lighting, background, or object appearance changes on the same robot.	A different robot changes appearance together with reach, grasp, or control conditions.
Task adaptation	The same robot learns a new procedure or task.	The same goal requires a different grasp or motion on another body.
Sim-to-real	The same robot is transferred from simulation to hardware and only the physical domain changes.	Sim-to-real transfer is combined with transfer to a different robot embodiment, together with changes in the control interface or contact conditions.
System integration	General wiring, configuration, and implementation.	Calibration, control connection, and safety checks needed to run a shared policy or representation on a new robot.
Action-space mismatch	Only the output format differs.	The same action representation produces different contact or recovery problems on another robot.
Transfer learning	General reuse across domains or tasks.	Transferred structure must be made executable again under the body and control system of the target robot.

Appendix

Table 4: Relation between adjacent concepts and the embodiment gap.

Stage	What is primarily shared	Interpretation for placement
Semantic / Task	Goals, instructions, plans	The desired outcome is shared, while the physical action remains unspecified.
Perception / Affordance	Cues about objects and places for action	The target or location is known, but it must be converted into motion on the target robot.
Object Interaction	How manipulation should change an object or scene	The desired change is shared, while the action and contact that produce it depend on the embodiment.
Action / Skill	Latent actions, action tokens, skills, trajectories	Action structure close to execution is shared, but it must be connected to the target robot’s action space.
Morphology-Aware Sharing	Body graphs, kinematics, body-specific conditions	Body differences are represented inside the policy to generate actions closer to the target robot.

Appendix

Table 5: Stages of shared structure used on the horizontal axis.

Stage	Work that primarily remains
Skill / API Connection	Connect plans or meaning to skills, functions, or planners that the target robot can actually call.
Calibration / Control Interface Alignment	Convert model outputs into commands that have the intended meaning in the action space and controller of the target robot.
Contact / Force Execution	Turn a spatially valid motion into physical contact, such as grasping or insertion, and correct deviations during contact.
Safety / Recovery	Stop safely after failures or disturbances and continue through retrying or recovery.

Appendix

Table 6: Stages at which work remains on the target robot, used on the vertical axis.

System / method group	Primary shared structure	Where work remains	Research direction	Placement rationale
SayCan	Semantic / Task	Skill / API Connection	Semantics and perception	Shares language-level task structure; the target robot still needs to connect each candidate to an executable skill.
RT-2 / PaLM-E	Semantic / Task	Skill / API Connection	Semantics and perception	Shares semantic and vision-language knowledge from web data; the target robot still needs to connect that knowledge to available actions or skills.
CLIPort / PerAct	Perception / Affordance	Calibration / Control Interface Alignment	Semantics and perception	Shares spatial cues about objects and places for action; the target robot still needs to align the output with motion generation and its control interface.
RT-Affordance / RoboPoint / VoxPoser	Perception / Affordance	Contact / Force Execution	Semantics and perception	Shares affordances or spatial cues for interaction; the target robot still needs to realize those cues as a stable grasp or contact.
VPP / FlowDreamer	Perception / Affordance	Calibration / Control Interface Alignment	Semantics and perception / cross-embodiment correspondence	Shares visual temporal change predicted from video or flow; the target robot still needs to convert it into executable actions and align them with its controller.
DROID / LeRobot / RoboMIND	Action / Skill	Calibration / Control Interface Alignment	Robot data and interfaces	Reuses robot experience and trajectories in common formats; the target robot still needs to map the action representation in the data to its own control system.

Appendix

Table 7: Adjudicated placements and rationales for representative method groups in Figure 2.

Item	Information to report when available
Target-robot data	Demonstrations, adaptation trials, failure and recovery data
Model updates	Frozen components, adapter/LoRA, decoder, full fine-tuning
Calibration and setup	Cameras, coordinate frames, gripper, control rate, workspace
Real-robot operation	Robot time, resets, reruns
Safety, intervention, and recovery	Stops, unsafe contact, manual or autonomous recovery
Evaluation rollouts	Held-out rollouts, tasks, seeds, success criteria

Appendix

Table 8: Compact reporting checklist.

Item	Information identifiable from public materials	Reading through the embodiment gap
Source robot(s) / source setting	Pretraining on Open X-Embodiment and reported out-of-box settings	Starting point of the shared VLA backbone and action tokens
Target robot	Fine-tuning setting on Franka	Conditions for execution on the target robot
Shared structure	7B VLA backbone and tokenized actions	Components reused across settings
Modified components	Full fine-tuning or LoRA adaptation	Components changed for the target robot
Target-robot data	10–150 demonstrations per Franka task across seven tasks	Target-robot experience used for adaptation
Model updates	Full fine-tuning or LoRA; rank-32 LoRA trains a subset of parameters	Update cost is partly visible

Appendix

Table 9: Worked report-card example for OpenVLA.

Observed failure	Primary cause	Secondary cause
Object flow is correct, but the gripper slips	Execution on the target robot	Correspondence across embodiments
The predicted interaction point is outside the reachable workspace	Correspondence across embodiments	Execution on the target robot
The action token does not match the control rate	Robot data and interfaces	Execution on the target robot
The task plan is correct, but no usable skill is available	Semantics and perception	Skill / API connection

Appendix

Table 10: Examples of organizing failure causes.

Explore similar work

Robots Need More than VLA and World Models
Jun 4, 2026Elis Karcini, Faisal Mehrban, Quang Nguyen +6Robot Foundation Models Cross-Embodiment Robot Learning
Stanford University · Istituto Italiano di Tecnologia · ETH Zurich +2
Data Pyramid for Embodied Manipulation: A Survey
Jul 27, 2026Yifan Ye, Yankai Fu, Yaoxu Lv +26Robotic Data Collection Robot Foundation Models
PKU · NTU · HKUST +8
Foundation Models in Robotics: A Comprehensive Review of Methods, Models, Datasets, Challenges and Future Research Directions
Apr 16, 2026Aggelos Psiris, Vasileios Argyriou, Evangelos K. Markakis +4Large Language Model-Based Robot Planning Robot Foundation Models
Department of Informatics and Telematics, Harokopio University of Athens, Athens, Greece · Department of Networks and Digital Media, Kingston University, London, United Kingdom · Department of Electrical and Computer Engineering, Hellenic Mediterranean University, Heraklion, Greece +4

Jun 4, 2026J/K move · Enter open · S save

Robots Need More than VLA and World Models

Elis Karcini, Faisal Mehrban, Quang Nguyen, Mac Schwager+5

Stanford University · Istituto Italiano di Tecnologia · ETH Zurich+2

Robot Foundation Models Cross-Embodiment Robot Learning Embodied World Models Robot Learning

Open paper AskPDF arXiv

Generalist robot intelligence is often framed as a policy-scaling problem: collect more robot demonstrations, train larger Vision-Language-Action (VLA) models, and expect broader generalisation. In this position paper, we argue that this framing is incomplete. The central bottleneck is not only policy learning, but the absence of mechanisms that convert the world's abundant unstructured behavioural data into grounded robot supervision. Human motion, internet video, simulation rollouts, and interactive demonstrations contain rich information about tasks, goals, contacts, failures, and physical constraints, yet most of this information is not directly usable by robot policies because it lacks embodiment-specific action labels, task semantics, and reward structure. We identify four missing components for the next generation of robotics: data interfaces for autolabelling unstructured behaviour, embodiment interfaces for retargeting human motion to robot actions, world-model interfaces for physics-grounded 3D reasoning, and reward interfaces for inferring task progress and success from video and language. We survey recent progress in robot foundation models, cross-embodiment datasets, learning from video, world models, and reward modelling, and propose a research agenda for building robotics systems that can learn not only from robot demonstrations, but from the broader physical world.

Abstract

Explore similar work

Robots Need More than VLA and World Models

Data Pyramid for Embodied Manipulation: A Survey

Foundation Models in Robotics: A Comprehensive Review of Methods, Models, Datasets, Challenges and Future Research Directions