Unified Navigation Representation (UNR)

• Unified Navigation Representation (UNR) is a design principle that integrates navigation variables across perception, mapping, planning, prediction, and control into one cohesive state.
• It leverages semantic, spatial, and temporal fusion—via approaches like latent feature streams and structured scene graphs—to improve long-horizon stability and task performance.
• Various implementations of UNR, including predictive shaping and algebraic encodings, address domain-specific trade-offs and enhance robustness in complex navigation tasks.

Unified Navigation Representation (UNR) denotes a design principle in which navigation-relevant variables that are usually separated across perception, mapping, planning, prediction, and control are encoded within a single representation, model, or state object. In the cited literature, this principle appears in several distinct forms: fused semantic–spatial features for streaming Vision-Language Navigation (VLN), shared latent spaces that align language, geometry, and future cues, unified tokenizations for heterogeneous embodied tasks, structured scene graphs and maps, continuous guiding fields, and compact algebraic encodings of inertial state (Fan et al., 4 Mar 2026). Across these formulations, the recurrent goal is to replace fragmented pipelines with representations that preserve consistency across modalities, time, and action.

1. Conceptual scope and main formulations

The surveyed papers use the term UNR for multiple representational regimes rather than for one canonical data structure. In embodied VLN, UNR often means a latent feature space that jointly encodes semantics, geometry, memory, and action intent. In world-model formulations, it becomes a shared spatio-temporal state for perception, generation, and control. In map-centric systems, it appears as a graph or dense scene representation that couples topology with semantics. In inertial and control-theoretic work, it becomes an algebraic or field-based state that evolves under a single differential law (Wang et al., 4 Aug 2025).

Family	Unified object	Unified elements
Streaming VLN	Fused latent or token stream	semantic features, spatial features, instruction context, action history
World-model navigation	Shared latent world state	perception, multimodal action prediction, future generation, memory
Structured scene memory	Graph or dense map	topology, regions, objects, geometry, photometry, semantics
Algebraic/control formulations	Quaternion-like state or guiding field	attitude, velocity, position, path geometry, obstacle awareness, feedback

This diversity is not incidental. A common misconception is that UNR must be an end-to-end neural embedding. The literature instead includes token-sequence formulations such as Uni-NaVid and OmniNav, explicit scene graphs such as OVER-NAV and USS-Nav, Gaussian maps for autonomy in unstructured environments, and compact mathematical states such as the trident quaternion and the guiding vector field (Zhang et al., 2024). This suggests that “unified” refers less to a fixed datatype than to the elimination of representational fragmentation.

2. Semantic–spatial fusion in embodied visual navigation

A central contemporary interpretation of UNR is semantic–spatial fusion for VLN. PROSPECT defines a unified streaming navigation agent in which a frozen 2D foundation vision-LLM, SigLIP, produces semantic features $\mathbf{F}^{2\text{D}_t}$, and a frozen 3D streaming encoder, CUT3R, produces absolute-scale 3D spatial features $\mathbf{F}^{3\text{D}_t}$. Cross-attention fuses these into $\mathbf{F}^{\mathrm{fuse}_t}$, which is then mapped into the LLM embedding space and concatenated with instruction and historical context as a stream. The policy is written as $a_t = \mathrm{VLA}(I, \textrm{Stream}_{0:t})$, with $\textrm{Stream}_{0:t}$ incorporating the latest observations, sliding-window context, and long-term memory (Fan et al., 4 Mar 2026).

MonoDream uses the term UNR for a compact, shared latent space that jointly aligns visual semantics, global layout, depth/geometric cues, future cues, and language-grounded action intent from monocular input. Its representation is built from language embeddings $E_{\text{text}}$, visual embeddings $E_{\text{vis}}$, and a VLM backbone output $h_t = \text{MonoDream-Backbone}(S_t)$, where $S_t = [E_{\text{text}}, E_{\text{vis}}]$. The key claim is that a monocular agent can internally “dream” global and future spatial context that would otherwise be available to panoramic or depth-equipped agents (Wang et al., 4 Aug 2025).

STRNet adopts a different but related formulation: a unified spatio-temporal representation built before the policy head. Its spatio-temporal fusion module performs spatial graph reasoning within each frame and models temporal dynamics through a hybrid temporal shift module combined with multi-resolution difference-aware convolution. The output representation $\mathbf{c}_T$ is used for both a diffusion-based policy head and a temporal distance head. Ablation results reported in the summary are unusually direct: removing either the spatial or temporal fusion module decreases success rate from 98% to 88% or 38%, and removing both yields 28% (Ren et al., 3 Apr 2026).

These systems converge on the same structural idea. Semantic relevance alone is insufficient for robust navigation, and geometric structure alone is insufficient for language-grounded control. The UNR therefore becomes a fusion space in which instruction-conditioned semantics, persistent spatial structure, and temporally evolving context are jointly available to the policy.

3. Predictive, generative, and memory-augmented UNR

A second major line of work treats UNR not merely as a fused perceptual state but as a predictive or generative latent space. PROSPECT introduces learnable query tokens $\mathbf{F}^{3\text{D}_t}$0 and $\mathbf{F}^{3\text{D}_t}$1 that are appended to each turn during training and are supervised to predict next-step SigLIP and CUT3R latent features rather than pixels or explicit modalities. The losses are a cosine-distance semantic loss,

$\mathbf{F}^{3\text{D}_t}$2

and a spatial mean-squared error,

$\mathbf{F}^{3\text{D}_t}$3

combined with navigation cross-entropy as

$\mathbf{F}^{3\text{D}_t}$4

The predictive branch is removed after training, so the paper emphasizes predictive shaping without inference-time cost (Fan et al., 4 Mar 2026).

MonoDream pursues a parallel idea through Latent Panoramic Dreaming (LPD). The model uses only monocular RGB at inference, but during training it is supervised to match the latent features of panoramic RGB and depth observations at both the current and future step. The feature loss is

$\mathbf{F}^{3\text{D}_t}$5

with $\mathbf{F}^{3\text{D}_t}$6. This loss is combined with action prediction and instruction reasoning, again yielding a training-only auxiliary pathway that enriches the navigation state (Wang et al., 4 Aug 2025).

NavWM extends the predictive view into a full unified navigation world model. It uses latent world tokens to distill geometric and semantic priors, a shared State Space Model backbone, anchor-based multimodal trajectory forecasting, and a Conditional Diffusion Transformer for future visual generation. Its action objective is

$\mathbf{F}^{3\text{D}_t}$7

while visual generation uses a flow-matching loss

$\mathbf{F}^{3\text{D}_t}$8

The reported offline results include PSNR $\mathbf{F}^{3\text{D}_t}$9 versus $\mathbf{F}^{\mathrm{fuse}_t}$0, LPIPS $\mathbf{F}^{\mathrm{fuse}_t}$1 versus $\mathbf{F}^{\mathrm{fuse}_t}$2, and navigation SR $\mathbf{F}^{\mathrm{fuse}_t}$3 versus $\mathbf{F}^{\mathrm{fuse}_t}$4 in seen environments and $\mathbf{F}^{\mathrm{fuse}_t}$5 versus $\mathbf{F}^{\mathrm{fuse}_t}$6 in unseen zero-shot environments (Mei et al., 23 Jun 2026).

UniWM likewise unifies planning and imagination inside a single multimodal autoregressive backbone. Its core step is

$\mathbf{F}^{\mathrm{fuse}_t}$7

with hierarchical memory comprising intra-step and cross-step caches fused by cosine-similarity gating and temporal decay. The paper reports success-rate gains up to 30% across Go Stanford, ReCon, SCAND, and HuRoN, and zero-shot improvement on TartanDrive from $\mathbf{F}^{\mathrm{fuse}_t}$8 to $\mathbf{F}^{\mathrm{fuse}_t}$9 SR when full memory is used (Dong et al., 9 Oct 2025).

P3Nav uses the language of a unified framework rather than an explicitly named standalone UNR, but its effective state is a multimodal embedding that fuses current 2D and 3D visual tokens, position-encoded historical observations, and linguistic input. The distinctive mechanism is Adaptive 3D-aware History Sampling, which retains only non-redundant historical frames under a distance threshold $a_t = \mathrm{VLA}(I, \textrm{Stream}_{0:t})$0, then augments them with positional encoding. It jointly trains navigation and embodied question answering through

$a_t = \mathrm{VLA}(I, \textrm{Stream}_{0:t})$1

and reports a 75\% success rate on the $a_t = \mathrm{VLA}(I, \textrm{Stream}_{0:t})$2-$a_t = \mathrm{VLA}(I, \textrm{Stream}_{0:t})$3 benchmark (Zhong et al., 24 Mar 2025).

4. Structured graphs, maps, and scene memories

Another substantial tradition treats UNR as explicit structure. OVER-NAV introduces the omnigraph, a semantic, ego-centric topological map whose nodes are viewpoints and whose edges encode connectivity discovered during navigation. The omnigraph stores open-vocabulary object detections linked to instruction-derived keywords, and its fusion module combines language embeddings, heading embeddings, and distance embeddings as

$a_t = \mathrm{VLA}(I, \textrm{Stream}_{0:t})$4

The framework is designed to support both discrete and continuous environments under a unified graph-based memory (Zhao et al., 2024).

USS-Nav defines a hierarchical Unified Spatio-Semantic scene graph with four layers: a local occupancy grid $a_t = \mathrm{VLA}(I, \textrm{Stream}_{0:t})$5, a global Spatial Connectivity Graph $a_t = \mathrm{VLA}(I, \textrm{Stream}_{0:t})$6, region clusters $a_t = \mathrm{VLA}(I, \textrm{Stream}_{0:t})$7, and object nodes $a_t = \mathrm{VLA}(I, \textrm{Stream}_{0:t})$8. Polyhedral expansion incrementally approximates free space, Leiden clustering partitions the graph into semantic regions, and open-vocabulary objects are anchored by combined geometric and semantic similarity, including

$a_t = \mathrm{VLA}(I, \textrm{Stream}_{0:t})$9

The paper reports 15 Hz real-time update frequency on a resource-constrained platform and substantial improvements in SPL (Gai et al., 31 Jan 2026).

“Gaussian Splatting as a Unified Representation for Autonomy in Unstructured Environments” advances a dense-map interpretation of UNR. Here each 3D Gaussian can carry position, orientation, scale, color, opacity, and semantic features, and uncertainty-driven exploration is defined by

$\textrm{Stream}_{0:t}$0

The paper reports RT-GuIDE map reconstruction with PSNR $\textrm{Stream}_{0:t}$1 versus $\textrm{Stream}_{0:t}$2 and SSIM $\textrm{Stream}_{0:t}$3 versus $\textrm{Stream}_{0:t}$4, while ATLAS uses 8–10 GB on HM3D scene 00824 with 2–5 m submaps (Ong et al., 17 May 2025).

An earlier visual-navigation formulation unified map-based reasoning with landmark-based execution. “Unifying Map and Landmark Based Representations for Visual Navigation” learns an allocentric map $\textrm{Stream}_{0:t}$5, a path planner, a feature-synthesis engine that predicts expected features along the planned path, and a closed-loop controller that follows the resulting path signature

$\textrm{Stream}_{0:t}$6

The reported success rates rise from about 27\% for open-loop execution and about 43\% for GRU-based policies with all visual memories along the path to about 80.5\% for the proposed architecture (Gupta et al., 2017).

A non-robotic but conceptually related precursor is the partitioning of a navigation domain into independent segments, each represented by a reference image $\textrm{Stream}_{0:t}$7 and auxiliary information $\textrm{Stream}_{0:t}$8. The segmenting rule

$\textrm{Stream}_{0:t}$9

shows an early form of unification: arbitrary viewpoints inside a segment are reconstructed from one segment-level representation, with additional data required only on segment transitions (Maugey et al., 2012).

5. Cross-task and cross-embodiment unification

Several papers broaden UNR from representation sharing within one task to common interfaces across many tasks. OmniNav converts instructions, object categories, point targets, and exploration directives into standardized multimodal tokens. Its fast system predicts continuous-space waypoints

$E_{\text{text}}$0

while a slow system reasons over frontiers, memory, and semantics. The paper reports state-of-the-art success rates across instruct-goal, object-goal, point-goal, and frontier-based exploration, including +4.4\%/4.3\% SR over previous SOTA on R2R-CE and RxR-CE, +20 pt SR/SPL improvement on open-vocabulary OVON with the slow system, mean average orientation error $E_{\text{text}}$1 versus $E_{\text{text}}$2 on Citywalker, and real-world deployment at control frequencies up to 5 Hz (Xue et al., 30 Sep 2025).

Uni-NaVid unifies VLN, ObjectNav, Embodied Question Answering, and Human Following at the token level. Ego-centric RGB video is encoded into observation tokens, compressed via online token merging into current, short-term, and long-term memory tokens, then concatenated with task markers such as <NAV> and natural-language instructions. The model is trained with next-token prediction on 3.6 million navigation data samples from four essential navigation sub-tasks, and it predicts the next $E_{\text{text}}$3 actions at once for asynchronous execution. The reported real-world numbers include inference time about 0.2 s per step and 5 Hz speed, along with 92\% success on simple and 84\% on complex real-world robot tasks (Zhang et al., 2024).

NaviMaster extends the unification principle across digital and physical environments. It models both GUI navigation and embodied navigation as a Markov Decision Process, standardizes actions via explicit visual targets such as [CLICK (x, y)] and [MOVETO (x, y)], and optimizes a single policy with Group Relative Policy Optimization and a distance-aware dense reward

$E_{\text{text}}$4

The summary reports higher embodied-navigation Success Rate and SPL than the Qwen2.5VL-7B base, including 33.2\% SR versus 27.2\% (Luo et al., 4 Aug 2025).

Uni-LaViRA presents a training-free, zero-shot formulation in which navigation reduces to a Language-Vision-Robot Actions Translation. The factorization

$E_{\text{text}}$5

splits each step into a language action, a vision action, and an embodiment-specific robot action. Two agent-loop mechanisms support long-horizon behavior: TODO List Memory (TDM), which rewrites a structured checklist of pending sub-goals at every step, and Second Chance Backtrack (SCB), which rolls the robot back to the pre-error state and conditions replanning on the failed sub-trajectory. The reported zero-shot results are 60.7\% SR on VLN-CE R2R, 51.3\% on VLN-CE RxR, 77.7\% on HM3D-v2, 60.0\% on HM3D-OVON, 54.7\% on MP3D-EQA, and 40.0\% on OpenUAV (Ding et al., 26 May 2026).

Taken together, these works define a strong version of UNR: not only a shared internal state, but a shared input–output grammar across heterogeneous tasks, environments, and embodiments.

6. Algebraic and control-theoretic unified navigation states

Outside embodied VLN, UNR also appears as a mathematically compact state representation. The trident quaternion framework introduces

$E_{\text{text}}$6

where $E_{\text{text}}$7, $E_{\text{text}}$8, and $E_{\text{text}}$9 encode attitude, velocity, and position, and $E_{\text{vis}}$0. The paper states that dual-quaternion formulations required three dual-quaternion ODEs totaling 24 states for 9 degrees of freedom, whereas the trident quaternion yields one 9-dimensional equation. It further reports that the tqFIter method outperforms the traditional two-sample algorithm by about 8 orders of magnitude in computation error for attitude, velocity, and position, with sensor data sampled at 100 Hz (Ouyang et al., 2021).

SwordRiding proposes a closed-loop UNR for quadrotors based on a guiding vector field synthesized from a Euclidean Signed Distance Field and a B-spline reference path. Its field is

$E_{\text{vis}}$1

where $E_{\text{vis}}$2 is the tangent estimated from discrete path points and $E_{\text{vis}}$3 is the normal derived from the distance field gradient. The representation is “unified” because environmental mapping, path following, and feedback recovery are encoded in the same geometric field. The reported advantage is improved robustness against external disturbances and superior real-time performance relative to open-loop planners and conventional GVF methods (Liu et al., 27 Nov 2025).

These two examples show that UNR is not restricted to multimodal deep learning. It can also mean a reduction of multiple navigation subsystems to one compact state evolution law or one continuous field representation.

7. Empirical patterns, misconceptions, and unresolved constraints

The empirical record summarized across these papers shows several recurring benefits. First, predictive supervision in latent space frequently improves navigation without test-time overhead. PROSPECT explicitly removes its predictive branch after training, and MonoDream applies LPD only during training while using only monocular RGB at inference (Fan et al., 4 Mar 2026). Second, unified world models and memory mechanisms are repeatedly linked to long-horizon stability: NavWM attributes gains to latent world reasoning plus multimodal planning, while UniWM reports that joint action–observation prediction with both intra- and cross-step memory outperforms no-memory variants (Mei et al., 23 Jun 2026).

A second recurring pattern is that explicit structure remains competitive with end-to-end latent modeling. OVER-NAV’s omnigraph, USS-Nav’s hierarchical scene graph, and Gaussian-splatting maps all maintain explicit spatial organization and semantic attachment rather than collapsing all information into one opaque embedding (Zhao et al., 2024). This suggests that the opposition between “neural UNR” and “structured UNR” is misleading; the literature contains both, and several systems combine them.

Several limitations are equally consistent. Gaussian splatting papers identify GPU memory as the limiting factor even with submapping, note that odometry alignment and LiDAR-camera misalignment can degrade map quality, and emphasize that Gaussians do not directly encode unobserved/free/occupied space (Ong et al., 17 May 2025). USS-Nav is motivated by the computational tension between high-level semantic reasoning and limited onboard resources on UAVs, which is why its LLM is called asynchronously and sparingly (Gai et al., 31 Jan 2026). In interactive multiview navigation, segment boundary transitions still require new server requests, and optimality depends on scene innovation and view-popularity modeling (Maugey et al., 2012).

A final misconception is that unification automatically implies one universal benchmark winner. The papers instead show domain-specific tradeoffs. P3Nav reaches 75\% success rate on $E_{\text{vis}}$4-$E_{\text{vis}}$5 but exhibits a SEL tradeoff because trajectories are longer though more successful (Zhong et al., 24 Mar 2025). PROSPECT reports especially large gains on the more challenging long-horizon RxR setting and stronger real-robot robustness under diverse lighting, including 20/30 successful missions in office/bright versus 12/30 for StreamVLN and 7/30 for NaVid (Fan et al., 4 Mar 2026). Uni-LaViRA, by contrast, shows that structural factorization plus agentic memory and backtracking can match or surpass trained foundation models on several benchmarks without any training effort (Ding et al., 26 May 2026).

The overall picture is therefore heterogeneous but coherent. UNR is best understood as a family of techniques for enforcing shared state across navigation-relevant factors that would otherwise be separated: semantics and geometry, current perception and future prediction, local control and global planning, discrete and continuous environments, or even attitude, velocity, and position. The surveyed literature suggests that progress in navigation increasingly depends on how these factors are unified, not only on how each component is optimized in isolation.