Cross-Embodiment Transfer in Vision-Based Robot Navigation: Challenges, Methods, and Open Problems

1. Introduction

Vision-based navigation has moved from single-task policies toward models trained on heterogeneous trajectories, large simulation corpora, and multiple robot platforms. GNM, ViNT, and NoMaD show that data from several mobile robots can be aggregated through shared goal-reaching objectives and standardized interfaces, yielding policies that are more reusable than platform-specific models [1]–[3]. More recent systems explicitly vary or encode embodiment: RING randomizes body and camera configurations in simulation, NavFoM inserts embodiment and camera identifiers into a unified navigation model, and EA-Nav conditions imitation learning on embodiment geometry [4]–[6].

These advances expose a central conceptual difficulty. An embodiment change is not a single shift in appearance. A taller camera observes different occlusion boundaries; a wider body changes which gaps are traversable; a legged robot and a differential-drive base interpret the same waypoint through different dynamics; and an error that is harmless for a compact platform can cause collision for a larger one. Consequently, a policy that maps images directly to actions can appear embodiment-agnostic while still hiding body-specific assumptions in its training distribution, action labels, or collision model.

This article develops a structured account of cross-embodiment transfer around one question: which parts of the navigation problem should be shared across robots, and which parts must remain explicitly conditioned on the target body? The answer leads to a factorized design principle. Scene representation should be as environment-centric as possible; embodiment parameters should be explicit when they affect observability or safety; collision and kinematic feasibility should be checked before low-level execution; and sim-to-real adaptation should be evaluated independently from cross-body generalization.

The article makes three contributions. First, it decomposes the embodiment gap into sensing, geometry, dynamics, action-interface, and reality components. Second, it organizes current methods by the mechanism through which they share knowledge across robots rather than by network architecture alone. Third, it proposes an evaluation matrix designed to reveal where transfer succeeds and where it merely benefits from correlated test conditions.

2. Problem Formulation: Transfer Is a Coupled Shift

Let the environment state at time t be s_t, the task or local goal be g_t, and the target embodiment be described by e. The descriptor may contain camera pose and intrinsics, collision dimensions, motion limits, controller type, and sensor availability. The robot receives an embodiment-dependent observation o_t^e and produces an action a_t^e.

Here, C denotes collision and feasibility constraints. Cross-embodiment transfer asks whether a policy trained on a set of source embodiments can maintain task performance and safety on a target embodiment that differs in one or more components of e, ideally without target-specific retraining.

This decomposition is important because methods often solve only a subset of the gaps. A model trained on many robots may learn a robust shared visual prior while still relying on a fixed collision radius. Conversely, a geometry-aware planner may transfer across body sizes while assuming the same camera modality and low-level controller. Claims of cross-embodiment performance should therefore state which components of e vary and which remain fixed.

3. A Taxonomy of Cross-Embodiment Methods

3.1. Heterogeneous-data generalist policies

The first family treats embodiment diversity as a data-sharing problem. GNM combines trajectories from six robots by using temporal context and a standardized action representation, demonstrating that heterogeneous data can produce a navigation model more robust than single-dataset specialists [1]. ViNT extends this idea through a Transformer trained with a general goal-reaching objective and supports adaptation to new task interfaces [2]. NoMaD unifies goal-directed navigation and exploration with a diffusion decoder trained on multi-robot data [3].

The strength of this family is scale: every additional robot can contribute experience to a common model. Its limitation is that diversity in the dataset does not automatically expose the causal variables responsible for safety. If body geometry is only implicitly correlated with visual appearance or action labels, the model may fail on target embodiments outside the training range. Generalist policies therefore benefit from explicit metadata, standardized actions, or a downstream safety layer.

3.2. Embodiment randomization and explicit conditioning

The second family makes the embodiment variable visible during training. RING randomizes body size, rotation pivot, and camera configuration in simulation to train a policy that generalizes across indoor robot platforms [4]. NavFoM scales cross-embodiment learning across quadrupeds, drones, wheeled robots, and vehicles by using identifiers for camera views and temporal context within a unified model [5]. EA-Nav directly addresses observation-action ambiguity by adding embodiment geometry as conditional tokens and by training risk-aware correction on augmented high-risk trajectories [6].

These methods support a more precise interpretation of transfer: the policy learns a family of behaviors indexed by e, rather than pretending that body differences are irrelevant. The design challenge is descriptor completeness. A compact vector that contains width and camera height may be sufficient for local obstacle avoidance but insufficient for platforms whose braking distance, gait phase, sensor latency, or articulated shape dominates safety.

3.3. Geometric abstraction and configuration-driven planning

The third family inserts an explicit geometric interface between vision and control. CeRLP converts heterogeneous visual observations into a unified geometric representation, including scale-corrected depth and height-adaptive scan abstraction, before local planning [7]. AgniNav uses a shared four-parameter collision envelope: collision-relevant height conditions monocular image-to-scan prediction, while front length, rear length, and half width configure collision checking in the local planner [8].

This factorization has a strong safety interpretation. Perception estimates environment structure relative to a declared body envelope; planning decides which actions are feasible for that envelope; and low-level control executes a platform-appropriate command. The interface is narrower than an end-to-end policy, but it makes transfer failures easier to diagnose. A collision can be attributed to perception error, incorrect envelope parameters, planner approximation, or controller tracking rather than to an opaque image-to-action mapping.

3.4. Sim-to-real transfer as an independent axis

Cross-body generalization and sim-to-real transfer should not be conflated. Habitat and Gibson made large-scale embodied training possible, but both also motivate careful analysis of what simulation preserves [9], [10]. Kadian et al. showed that simulator rankings can have weak correlation with real-world performance when agents exploit collision behavior or other simulator artifacts [11]. Anderson et al. found that transferring vision-and-language navigation becomes substantially harder when prior mapping assumptions are removed [12]. Truong et al. further showed that greater simulator fidelity can reduce transfer when slower training and inaccurate physics encourage overfitting [13]. Recent work using 3D feature fields illustrates a different strategy: explicitly reconstruct deployable visual representations for monocular robots rather than relying on rendering realism alone [14].

The implication is methodological: a system can generalize across simulated bodies yet fail on every physical robot, or transfer to one physical robot while remaining body-specific. Experiments should cross the two axes rather than report a single combined success rate.

4. Unifying Design Logic

The reviewed methods suggest a four-layer architecture for reliable cross-embodiment navigation.

1. Environment-centric representation. Map raw observations into geometry, traversability, semantic waypoints, or latent affordances that are useful across platforms. This layer should suppress body-specific nuisances without discarding information needed for safety.
2. Explicit embodiment interface. Provide camera, collision, and dynamics descriptors when they alter the meaning of an observation or goal. Learned tokens and measured physical parameters can coexist.
3. Feasibility and risk layer. Convert shared representation into body-aware constraints, candidate trajectories, or action rankings. Collision checking, safety critics, and control constraints belong here.
4. Embodiment-specific execution. Translate a standardized waypoint or velocity command into platform control while accounting for latency, tracking error, gait, braking, or balance.

This architecture explains when end-to-end learning and modular planning are complementary rather than competing. End-to-end models are effective for learning reusable visual affordances and priors from diverse data. Explicit geometric or constrained layers are valuable where physical dimensions determine hard feasibility. The best division is task-dependent: long-horizon semantic navigation may favor a generalist model at the top of the stack, while near-field collision avoidance requires a representation tied to measurable clearance.

A key distinction is between invariance and equivariance. Some features should remain invariant across embodiments, such as the identity of a doorway or the approximate location of free space. Other outputs should change predictably with embodiment, such as whether the doorway is traversable or how sharply the robot can turn. Treating all variation as something to remove can erase precisely the information required for safe behavior.

5. Evaluation: What Must Be Reported

Navigation evaluation already distinguishes success and path efficiency, but cross-embodiment studies require a stricter protocol. Standard guidance on embodied navigation emphasizes consistent task definitions and metrics such as success weighted by path length (SPL) [15]. For robots with different dynamics, success weighted by completion time (SCT) can reveal differences hidden by path length alone [16].

Results should also report the target descriptor used at deployment and how it was measured. A zero-shot result is less informative if target-specific calibration silently estimates a dense dynamics model, while a small amount of explicit geometry measurement may be a practical and reproducible assumption. The evaluation should distinguish reusable prior knowledge from target-specific information.

6. Failure Modes and Diagnostic Tests

6.1. Observation-action ambiguity

The same image and goal can require different actions for bodies of different width or turning capability. A policy without embodiment input must infer the body from incidental visual or temporal cues, which becomes unreliable on unseen platforms. A direct diagnostic is to hold the image and goal fixed while varying the declared embodiment and test whether predicted risk or action changes monotonically.

6.2. Camera-body inconsistency

Changing camera height alters visibility but does not necessarily change the collision envelope in the same way. Systems that use camera pose as a proxy for body geometry can fail when sensors are remounted or when tall sensors sit on compact bases. Camera parameters and collision parameters should therefore be represented separately unless their coupling is guaranteed by design.

6.3. Geometric success with dynamic failure

A path may clear the footprint while violating braking, balance, or tracking constraints. This is especially important when a common planner commands wheeled, quadruped, and humanoid systems. Evaluation should replay identical geometric paths through embodiment-specific controllers and measure overshoot, stopping distance, and tracking error.

6.4. Simulator exploitation

Agents may exploit permissive collision models, deterministic sensing, or unrealistic actuation. The simulator should be judged not only by visual fidelity but by whether model rankings and failure categories persist in reality [11], [13]. Perturbing collision response and sensor latency is often more revealing than increasing rendering quality.

6.5. Static safety margins

Fixed inflation margins create a predictable trade-off: conservative values reduce collision risk but cause detours and timeouts, while permissive values improve efficiency at the cost of near-boundary failures. Recent work on adaptive safety critics suggests that context-dependent clearance budgets can be learned, but these mechanisms must be constrained against margin collapse and validated on real hardware [17].

7. Open Problems

7.1. Online embodiment estimation

Most systems assume that embodiment descriptors are known. In practice, the effective envelope changes with payloads, manipulators, protective shells, gait posture, and sensor remounting. Future systems should estimate a posterior over embodiment parameters from calibration motion, proprioception, and near-miss observations, then propagate uncertainty into planning.

7.2. Continuous and articulated embodiment models

Width-height vectors are useful but incomplete. Humanoids and mobile manipulators change shape during motion. A general descriptor may need a time-varying swept volume, reachability set, or learned latent dynamics representation. The challenge is to retain enough structure for safety verification without losing the scalability of learned generalist policies.

7.3. Heterogeneous sensing without a privileged modality

Many transfer pipelines use depth during training or calibration even when deployment is monocular. A stronger formulation would allow arbitrary subsets of RGB, depth, LiDAR, event, and proprioceptive inputs while exposing uncertainty to the planner. Missing-modality robustness should be evaluated independently from body transfer.

7.4. Safety guarantees for learned perception

Collision checking is only as reliable as the geometric estimate supplied to it. Formal guarantees therefore require uncertainty bounds on perception, conservative reachable sets, or runtime monitors that can override the learned policy. A promising direction is to combine large learned proposal models with small verifiable safety filters.

7.5. Benchmarks that separate the gaps

Current benchmarks often vary scene, robot, and simulator simultaneously. This makes it difficult to attribute improvements. A useful cross-embodiment benchmark should include controlled pairs that vary one factor at a time, plus compositional tests that combine novel geometry, sensor, dynamics, and reality conditions.

7.6. Foundation models with explicit physical interfaces

Navigation foundation models provide semantic priors, long-horizon memory, and multi-task transfer, but safe deployment requires a physical contract between high-level reasoning and low-level feasibility. Future models should output uncertainty-aware subgoals or trajectory distributions together with the embodiment assumptions under which those outputs are valid.

8. Conclusion

Cross-embodiment visual navigation is not solved by making a policy blind to the robot. The transferable system must instead distinguish shared environmental structure from body-dependent observability, feasibility, and control. Heterogeneous-data policies contribute reusable priors; embodiment-conditioned models resolve observation-action ambiguity; geometric abstractions expose measurable safety constraints; and sim-to-real research clarifies which conclusions survive physical deployment. The resulting design principle is factorization with explicit interfaces: share what can be shared, condition what changes the meaning of perception, verify what determines safety, and evaluate each transfer gap separately.

References

1. D. Shah, A. Sridhar, A. Bhorkar, N. Hirose, and S. Levine. GNM: A General Navigation Model to Drive Any Robot. arXiv:2210.03370, 2022.
2. D. Shah, A. Sridhar, N. Dashora, K. Stachowicz, K. Black, N. Hirose, and S. Levine. ViNT: A Foundation Model for Visual Navigation. arXiv:2306.14846, 2023.
3. A. Sridhar, D. Shah, C. Glossop, and S. Levine. NoMaD: Goal Masked Diffusion Policies for Navigation and Exploration. arXiv:2310.07896, 2023.
4. A. Eftekhar et al. The One RING: a Robotic Indoor Navigation Generalist. arXiv:2412.14401, 2024.
5. J. Zhang et al. Embodied Navigation Foundation Model. arXiv:2509.12129, 2025.
6. J. Zhang, Y. Du, X. Guo, S. Sun, X. Liu, Y. Sun, G. Lu, W. Sui, and J. Li. EA-Nav: Learning Safe Visual Navigation Policies with Embodiment Awareness. arXiv:2607.19880, 2026.
7. H. Xi, M. Tan, X. Zhang, S. Cheng, S. Wang, Y. Gu, X. Shen, and W. Zhang. CeRLP: A Cross-embodiment Robot Local Planning Framework for Visual Navigation. arXiv:2603.19602, 2026.
8. T. Zang, S. Cheng, H. Huang, S. Wang, and W. Zhang. AgniNav: Configuration-Driven Cross-Embodiment Local Planning for Robot Navigation. arXiv:2606.10903, 2026.
9. M. Savva et al. Habitat: A Platform for Embodied AI Research. arXiv:1904.01201, 2019.
10. F. Xia, A. Zamir, Z.-Y. He, A. Sax, J. Malik, and S. Savarese. Gibson Env: Real-World Perception for Embodied Agents. arXiv:1808.10654, 2018.
11. A. Kadian et al. Sim2Real Predictivity: Does Evaluation in Simulation Predict Real-World Performance? arXiv:1912.06321, 2019.
12. P. Anderson, A. Shrivastava, J. Truong, A. Majumdar, D. Parikh, D. Batra, and S. Lee. Sim-to-Real Transfer for Vision-and-Language Navigation. Proceedings of Machine Learning Research, 155:671–681, 2021.
13. J. Truong, M. Rudolph, N. H. Yokoyama, S. Chernova, D. Batra, and A. Rai. Rethinking Sim2Real: Lower Fidelity Simulation Leads to Higher Sim2Real Transfer in Navigation. Proceedings of Machine Learning Research, 205:859–870, 2023.
14. Z. Wang, X. Li, J. Yang, Y. Liu, and S. Jiang. Sim-to-Real Transfer via 3D Feature Fields for Vision-and-Language Navigation. Proceedings of Machine Learning Research, 270:2982–2995, 2025.
15. P. Anderson et al. On Evaluation of Embodied Navigation Agents. arXiv:1807.06757, 2018.
16. N. Yokoyama, S. Ha, and D. Batra. Success Weighted by Completion Time: A Dynamics-Aware Evaluation Criteria for Embodied Navigation. arXiv:2103.08022, 2021.
17. J. Hu, S. Yuan, G. C. R. Bethala, A. Tzes, and Y. Fang. Learning Adaptive Safety Margins for Visual Navigation. arXiv:2607.18200, 2026.