NAVMOE: Adaptive Navigation via Mixture Experts

• NAVMOE is a conditional navigation framework that selects specialized expert modules based on scene, terrain, and contextual cues.
• It partitions heterogeneous navigation regimes to reduce interference among subskills while optimizing computational efficiency.
• Implementations span hybrid traversability estimates, sparse transformer experts, and state-adaptive routing in diverse navigation tasks.

Searching arXiv for recent NAVMOE-related papers to ground the article. Navigation via Mixture of Experts (NAVMOE) denotes a class of navigation systems in which expert selection is conditioned on the current scene, state, terrain, modality, or routing context rather than being fixed across all inputs. The term is explicit in "NavMoE: Hybrid Model- and Learning-based Traversability Estimation for Local Navigation via Mixture of Experts" (He et al., 16 Sep 2025), but the broader paradigm also appears in socially compliant robot navigation, instruction-guided visual navigation, autonomous-driving trajectory planning, and terrain-adaptive humanoid locomotion (Kawabata et al., 15 Dec 2025, Zhou et al., 2024, Xing et al., 8 Dec 2025, Zhu et al., 18 May 2025, Jiang et al., 13 Nov 2025, Ma et al., 3 Mar 2026). In these works, MoE is not treated merely as a scaling device; it is used to partition heterogeneous navigation regimes, reduce interference among subskills, and improve the performance–efficiency balance under real-time constraints.

1. Scope and conceptual boundaries

NAVMOE spans multiple levels of the navigation stack. In local navigation, the output can be a traversability map used by a primitive-based planner, as in NAVMOE’s hierarchical traversability estimator and local planner (He et al., 16 Sep 2025). In social navigation, the output can be a language-like next-step action instruction conditioned on image and conversational context, as in SocialNav-MoE (Kawabata et al., 15 Dec 2025). In autonomous driving, MoE can operate inside a trajectory-vocabulary scorer, a scene-adaptive planner decoder, or a prediction–planning interface (Xing et al., 8 Dec 2025, Zhu et al., 18 May 2025, Jiang et al., 13 Nov 2025). In embodied visual-language navigation, it can modulate action logits over candidate viewpoints through state-adaptive cross-modal experts (Zhou et al., 2024). In humanoid locomotion, it can adapt low-level control to mixed terrains by conditioning actor-critic experts on exteroceptive terrain structure (Ma et al., 3 Mar 2026).

This breadth matters because it clarifies that NAVMOE is a design principle rather than a single architecture. Some systems route over internal transformer sublayers, some over semantically named expert modules, some over black-box heterogeneous estimators, and some over actor-critic policies. A related but broader precursor is the LLM-enabled MoE framework for compositional control, which demonstrated expert composition on a $3 \times 7$ maze navigation task via natural-language task parsing and expert weighting (Du et al., 2024). Another adjacent line is dynamic routed multimodal transformers, where routing determines which existing layers or adapters are executed per example rather than which parallel experts are selected (Wu et al., 2024). These works suggest that “navigation via mixture of experts” includes both explicit mixtures of specialized navigators and routed computation graphs whose active path is itself navigation-conditioned.

A common misconception is that NAVMOE implies a single type of expert specialization, such as maneuver classes or terrain classes. The literature shows otherwise. Specialization may be explicit and label-driven, as in EMoE-Planner’s seven scenario experts (Zhu et al., 18 May 2025); implicit and state-adaptive, as in SAME (Zhou et al., 2024); semantically structured around environment, ego-state, and route intent, as in ExpertAD (Jiang et al., 13 Nov 2025); or hybrid across model-based and learning-based traversability estimators, as in NAVMOE (He et al., 16 Sep 2025).

2. Architectural forms of NAVMOE

Representative NAVMOE systems differ mainly in where the expert decomposition is inserted and what object is routed.

System	Expert locus	Routing granularity
NAVMOE (He et al., 16 Sep 2025)	Hybrid traversability estimators	Domain-level and pixel-level
SocialNav-MoE (Kawabata et al., 15 Dec 2025)	Sparse transformer FFN experts in a small VLM	Token-level
SAME (Zhou et al., 2024)	Cross-modal visual-query experts in DUET	Per-timestep state
TrajMoE / EMoE-Planner (Xing et al., 8 Dec 2025, Zhu et al., 18 May 2025)	Planner transformer experts	Scene-conditioned or token-conditioned
ExpertAD (Jiang et al., 13 Nov 2025)	Prediction/planning experts over environment, ego, navigation factors	Top-$k$ scene routing
CMoE (Ma et al., 3 Mar 2026)	Actor-critic experts for locomotion	Terrain-conditioned mixture

The architectural contrast between these systems is substantial. SocialNav-MoE does not sparsify every block; half of the transformer layers are replaced with sparse MoE layers and the remaining half stay as dense FFNs, explicitly to balance representation capacity and training stability in a scarce-data robotics setting (Kawabata et al., 15 Dec 2025). SAME does not use MoE primarily in FFNs; its best configuration places experts on the visual query projection inside cross-attention, which means expertization alters how observations query the instruction rather than merely changing per-token feed-forward processing (Zhou et al., 2024). EMoE-Planner uses an explicit scene router that predicts one of seven scenario categories and then reuses the same expert index across all decoder layers, yielding a semantically interpretable form of expert reuse (Zhu et al., 18 May 2025).

A second architectural axis is whether experts are homogeneous subnetworks or heterogeneous modules. In TrajMoE and SocialNav-MoE, experts replace standard FFN sublayers inside transformers (Xing et al., 8 Dec 2025, Kawabata et al., 15 Dec 2025). In ExpertAD, the experts are semantically differentiated modules: Tracking Expert, Mapping Expert, Velocity Expert, Yaw Expert, Acceleration Expert, Reference Point Expert, BEV Expert, and Command Expert (Jiang et al., 13 Nov 2025). In NAVMOE, the experts can be either classical model-based or learning-based traversability predictors, including Falco, Mask2Former, MaskFormer, and GA-Nav (He et al., 16 Sep 2025). This hybridization is especially important because it makes nondifferentiable geometric estimators first-class components of the MoE rather than auxiliary post-processors.

A third axis is whether routing controls representation learning or final decision formation. TrajMoE and EMoE-Planner insert MoE into the planner’s internal trajectory-scoring computation (Xing et al., 8 Dec 2025, Zhu et al., 18 May 2025). ExpertAD routes the prediction/planning interface after perception adaptation (Jiang et al., 13 Nov 2025). NAVMOE routes among traversability estimators before local planning (He et al., 16 Sep 2025). SAME routes cross-modal grounding during action selection itself (Zhou et al., 2024). CMoE routes actor and critic outputs in low-level locomotion control (Ma et al., 3 Mar 2026). The literature therefore supports a layered view of NAVMOE: expertization can occur at traversability estimation, multimodal reasoning, proposal scoring, trajectory planning, or control.

3. Routing mechanisms and conditional computation

Routing is the core technical differentiator of NAVMOE systems. SAME uses a standard sparse MoE form,

$$\mathcal{P}(\mathbf{x}_r) = \text{Softmax}(\mathcal{R}(\mathbf{x}_r)), \qquad \text{MoE}(\mathbf{x}, \mathbf{x}_r) = \sum_{i \in \mathcal{T}} \mathcal{P}(\mathbf{x}_r)_i \, f_i(\mathbf{x}),$$

but its routing feature is not a token identity or task label; it is a multimodal state representation built from pooled visual observations and the instruction CLS embedding (Zhou et al., 2024). This is important because the selected experts may change within the same episode as the embodied state changes. The paper’s routing ablations show that multimodal state routing outperforms task-wise routing, and adding task embeddings hurts, especially on REVERIE (Zhou et al., 2024).

SocialNav-MoE uses token-level top-$k$ routing over sparse FFN experts, but the empirical lesson is that stronger sparsity is often better than broader averaging. With top-$k=1$, sentence mover’s similarity rises from 0.473 with one expert to 0.523 with four experts; with top-$k=2$, gains saturate earlier, and with four experts performance degrades (Kawabata et al., 15 Dec 2025). The paper attributes this to data scarcity and noisy aggregation among partially trained experts. ExpertAD shows a related pattern: Top-4 routing outperforms Top-8 activation, indicating that activating all experts is not equivalent to useful specialization (Jiang et al., 13 Nov 2025). kNN-MoE, though not a navigation paper, reports that $K=1$ nearest-neighbor retrieval is consistently best, again suggesting that high-fidelity sparse correction can outperform broader neighborhood averaging (Lyu et al., 5 Jan 2026).

Some NAVMOE systems replace purely parametric routing with memory or hierarchy. kNN-MoE defines a confidence-aware interpolation,

$$a_{\text{final}}(x) = (1 - \lambda(x))\, a(x) + \lambda(x)\, a_{\text{mem}}(x),$$

where $a(x)$ is the frozen router output and $a_{\text{mem}}(x)$ is a retrieved routing proposal from similar prior cases (Lyu et al., 5 Jan 2026). This is not a navigation controller, but it provides a directly relevant routing principle: under distribution shift, expert assignment can be treated as analogy-based reuse of prior routing states rather than as a fixed classifier decision.

NAVMOE’s own traversability system adds a different kind of routing hierarchy. A top router predicts domain weights from global DINOv2 features, a lower router predicts pixel-wise terrain-expert weights from dense features, and a training-free lazy gating mechanism then orders experts by a compute-aware score and stops early when planner-cost uncertainty falls below a threshold (He et al., 16 Sep 2025). The stopping rule is planner-aware rather than purely router-aware, which is unusual: it terminates based on whether additional expert evaluations could still materially change the selected local path.

EMoE-Planner is at the opposite end of the routing spectrum. Its routing is explicit, supervised, and semantically labeled: the router predicts one of seven scene categories, and the selected expert is reused across decoder layers (Zhu et al., 18 May 2025). No load-balancing loss is used because the expert semantics are externally specified. This design trades latent flexibility for interpretability and direct per-scenario debugging.

4. Training objectives and optimization strategies

NAVMOE methods typically require training procedures that go beyond ordinary supervised learning. SocialNav-MoE uses a three-stage pipeline: supervised fine-tuning, reinforcement fine-tuning, and MoE fine-tuning. Its key technical addition is semantic similarity reward, defined through BERTScore-F1, to address the ambiguity of socially compliant action language (Kawabata et al., 15 Dec 2025). The paper reports that SSR is more effective than hard-level and character-level rewards, and that GSPO+SSR improves SMS from 0.527 under DPO to 0.551. This is a navigation-specific training point: when outputs are language-like high-level actions rather than deterministic control targets, semantically smooth rewards are better aligned with the task than exact-match objectives.

TrajMoE also uses reinforcement learning, but in a narrower role. GRPO is applied to fine-tune the scoring heads of a GTRS-style trajectory scorer rather than the entire planner, and rewards are based on deviation from ground-truth score (Xing et al., 8 Dec 2025). The paper’s own framing is that one-stage supervised training leaves the evaluation mechanism under-refined, whereas RL improves score calibration. This is distinct from policy-gradient trajectory optimization in the conventional control sense; it is RL-style refinement of a proposal-ranking module.

Several systems emphasize balancing or regularization to avoid expert collapse. SAME optimizes

$k$0

with $k$1, and reports that load balancing and sequential dataset sampling are important for stable multitask training (Zhou et al., 2024). TrajMoE introduces a balance loss based on expert importance and load coefficients of variation, reflecting the standard MoE concern that scenario-adaptive experts are useful only if the router actually distributes data across them (Xing et al., 8 Dec 2025). ExpertAD uses a switch-style balancing loss $k$2 to keep its eight experts from collapsing (Jiang et al., 13 Nov 2025).

Other papers highlight that training stability may require structural compromises. SocialNav-MoE alternates dense and sparse FFN layers rather than sparsifying the full stack (Kawabata et al., 15 Dec 2025). RoE, an adjacent dynamic-routing MLLM method, uses staged training—adapter warmup, router warmup, then full instruction tuning—and a difficulty-aware sparsity regularizer to align training and inference routing (Wu et al., 2024). While not a navigation paper, it provides a transferable recipe for routed multimodal policies in which shortcut inference should be stronger on easy samples than on hard ones.

NAVMOE’s traversability system faces a different problem: nondifferentiable experts and limited multimodal labels. Its solution is a two-stage strategy. First, routers are pretrained on 32,000 images from ADE20K, Mapillary Vistas, and RUGD using human-prior mappings from semantic classes to expert preferences. Second, routers and experts are fine-tuned on 3,000 multi-sensory samples by constructing weak routing supervision from whichever expert best matches ground-truth traversability at each pixel (He et al., 16 Sep 2025). This directly addresses a central practical barrier in hybrid NAVMOE systems: how to train gating when some experts are classical algorithms.

CMoE introduces yet another training pattern. It keeps a single-stage PPO framework but adds a SwAV-style contrastive objective aligning gate embeddings and terrain embeddings so that expert activation becomes terrain-discriminative (Ma et al., 3 Mar 2026). The paper does not specify a full joint loss weight for PPO plus contrastive terms, but its premise is explicit: vanilla MoE gating remains nearly uniform across terrains unless specialization pressure is added.

5. Empirical manifestations across navigation domains

The most explicit NAVMOE result is the traversability-based local navigation system in (He et al., 16 Sep 2025). Its full non-lazy variant reaches average $k$3 at 313.18 GFLOPS, while the final lazy-gated system reaches $k$4 at 40.22 GFLOPS. The abstract and conclusion highlight an 81.2% reduction in average computational cost with less than a 2% loss in path quality, while the body also reports 86.2% and the table implies roughly 87.2% relative to the non-lazy configuration. The reporting discrepancy does not change the qualitative result: planner-aware lazy gating retains nearly the same path quality while drastically reducing active compute (He et al., 16 Sep 2025).

In social robot navigation, SocialNav-MoE is explicitly optimized for the trade-off between socially compliant action quality and deployability. On augmented SNEI, it reports BERTScore-P 0.520, BERTScore-R 0.492, BERTScore-F1 0.506, SBERT-cos 0.671, and SMS 0.551, while using 5.74B parameters and achieving 1.709 FPS. GPT-4o is listed at about 200B parameters and 0.212 FPS, and Claude at about 175B and 0.087 FPS; SocialNav-MoE is therefore reported as 8.1× and 19.6× faster, respectively, while using only 2.9% and 3.3% of their parameter counts (Kawabata et al., 15 Dec 2025). The paper also notes that even its non-augmented version exceeds those baselines on SMS.

In autonomous driving, the evidence for NAVMOE is distributed across planner, decoder, and prediction layers. TrajMoE achieves 51.08 on the navsim ICCV benchmark and ranks third; on navhard, GTRS-Aug with V2-99 reaches 42.1 EPDMS, while TrajMoE-MoE with the same backbone reaches 43.8 (Xing et al., 8 Dec 2025). EMoE-Planner reports 93.06 / 90.67 / 80.12 on Val14 / Test14 / Test14-hard without post-processing, outperforming the listed learning-based baselines without post-processing on all three closed-loop splits (Zhu et al., 18 May 2025). ExpertAD reports reductions of average collision rates by up to 20% and inference latency by 25% compared to prior methods, with detailed gains such as Expert-UniAD improving Driving Score from 44.62 to 55.49 and reducing latency from 534 ms to 445 ms (Jiang et al., 13 Nov 2025). Across these works, MoE improves not only open-loop fit but also closed-loop robustness and runtime.

In multitask visual-language navigation, SAME demonstrates that one routed policy can address R2R, RxR-EN, REVERIE, ObjectNav-MP3D, CVDN, SOON, and R2R-CE within a shared framework (Zhou et al., 2024). The paper summarizes about a 3% average success-rate gain over the unified non-MoE baseline, with the clearest improvement on REVERIE, where SR increases from 40.39 to 45.67 in the three-task comparison. The gains are not uniform across tasks, which is itself informative: state-adaptive expertization is especially helpful where exploration and language grounding conflict within the same episode.

At the locomotion layer, CMoE shows that MoE can be effective below the planner proper. In simulation, success rates on mixed terrains rise from 0.605 to 0.767 on mix1 and from 0.662 to 0.747 on mix2 relative to vanilla MoE; real-world tests report traversal of continuous steps up to 20 cm and gaps up to 80 cm on the Unitree G1 humanoid (Ma et al., 3 Mar 2026). This does not constitute full navigation in the route-planning sense, but it demonstrates that terrain-adaptive expert routing can materially expand the feasible motion envelope of a navigation system’s control substrate.

6. Limitations, misconceptions, and unresolved issues

A recurring misconception is that MoE automatically yields interpretable specialization. The literature is more cautious. SocialNav-MoE does not provide router heatmaps or per-expert semantic-role analysis (Kawabata et al., 15 Dec 2025). SAME reports performance gains from state-adaptive routing but does not visualize expert specializations or vary the number of experts (Zhou et al., 2024). ExpertAD semantically names experts but does not deeply validate disentanglement or activation structure (Jiang et al., 13 Nov 2025). CMoE provides stronger qualitative evidence through t-SNE plots, gate trajectories, and an expert-removal test, but still does not report formal specialization metrics such as gate entropy or mutual information (Ma et al., 3 Mar 2026). The field therefore has stronger evidence for functional utility than for mechanistic understanding.

A second misconception is that more experts or denser activation must improve navigation. SocialNav-MoE, ExpertAD, and kNN-MoE all show the opposite in their respective settings: top-1 or top-4 sparse selection can outperform denser mixtures, and $k$5 retrieval can outperform larger neighborhoods (Kawabata et al., 15 Dec 2025, Jiang et al., 13 Nov 2025, Lyu et al., 5 Jan 2026). This suggests that low-data embodied regimes are particularly sensitive to noisy expert aggregation and under-specialized routing.

Evaluation remains another unresolved issue. Social navigation is measured by semantic similarity to textual references, yet the paper explicitly notes that socially acceptable alternatives may deviate from ground truth and still be valid (Kawabata et al., 15 Dec 2025). NAVMOE’s traversability work emphasizes path quality $k$6 rather than only map MSE, reflecting the fact that map fidelity matters mainly insofar as it changes the selected local path (He et al., 16 Sep 2025). TrajMoE refines score prediction rather than executing closed-loop control, whereas EMoE-Planner emphasizes closed-loop non-reactive simulation as the decisive metric (Xing et al., 8 Dec 2025, Zhu et al., 18 May 2025). These differences imply that “navigation quality” is not standardized across NAVMOE subfields.

Data scarcity and deployment gaps also remain prominent. SocialNav-MoE is trained on SNEI with 325 images, augmented to 530, and the paper explicitly frames many design choices as responses to low-data conditions (Kawabata et al., 15 Dec 2025). EMoE-Planner identifies U-turns and roundabouts as weak experts due to lower data frequency (Zhu et al., 18 May 2025). ExpertAD still struggles in Overtake and GiveWay scenarios requiring more human-like reasoning, and the paper states that rule-based fallback systems are still needed (Jiang et al., 13 Nov 2025). CMoE improves mixed-terrain traversal substantially, but mixed-terrain success rates remain below 0.8 in simulation (Ma et al., 3 Mar 2026). These are not failures of the NAVMOE idea; they indicate that expert routing alone does not eliminate sparse coverage, ambiguous supervision, or long-tail scenario difficulty.

Finally, routing robustness under distribution shift is an open concern. kNN-MoE is motivated precisely by the brittleness of frozen parametric routers and shows that retrieval-augmented routing helps most on high-perplexity inputs (Lyu et al., 5 Jan 2026). A plausible implication is that future NAVMOE systems may increasingly combine parametric gating with memory, retrieval, or confidence-aware fallback. The existing literature already points in that direction: explicit scene routing, planner-aware lazy activation, semantic rewards, training–inference alignment, and terrain–gate contrastive alignment all attempt to make expert selection more stable and more task-aware. Collectively, they suggest that NAVMOE is evolving from a simple sparse-computation motif into a broader methodology for conditional navigation intelligence.