Demand-Driven Navigation (DDN)

• Demand-Driven Navigation (DDN) is a family of paradigms that uses user demand signals—such as natural language instructions or demand forecasts—to condition navigation decisions.
• DDN integrates visual, linguistic, and spatial data across embodied AI, virtual worlds, and transportation to dynamically guide route planning and object selection.
• Key techniques include demand-conditioned attribute spaces, language-to-map grounding, and adaptive computation which enhance navigation success and efficiency in complex, real-world scenarios.

Demand-Driven Navigation (DDN) denotes several related navigation paradigms in which navigation is organized around a demand signal rather than a fixed route alone. In embodied AI, the demand is a natural-language need such as “I am thirsty,” and the agent must infer suitable goal objects or categories and navigate to any instance that satisfies the need (Wang et al., 2023, Long et al., 2024, Huang et al., 15 Jul 2025). In commercial virtual worlds, DDN is an on-demand, intent-driven approach to wayfinding delivered by a conversational spatial navigation agent that reacts to user requests in real time and adapts guidance to interests and context (Yanagawa et al., 5 Aug 2025). In transportation and dispatch, related work uses demand forecasts and uncertainty sets to guide vehicle repositioning and fleet sizing (Alsaleh et al., 2020, Miao et al., 2016), while a separate navigation literature uses the same acronym for “Data-Driven Navigation” in inertial navigation and sensor fusion (Klein, 2022). This suggests that DDN is best regarded as a family of demand-conditioned navigation formulations rather than a single standardized task.

1. Conceptual scope and terminological variation

The literature does not present a single canonical definition of DDN. Navigation Pixie contrasts DDN with fixed-route navigation and static wayfinding: fixed-route navigation constrains the user to a pre-scripted sequence, while DDN centers engagement on the user’s intent, expressed as free-form utterances, and couples intent inference, destination selection, feasible routing, and conversational guidance (Yanagawa et al., 5 Aug 2025). Foundational embodied work similarly defines DDN as replacing object names with a high-level user demand, thereby relaxing the Visual Object Navigation assumptions that the user can name the target object and that the named object exists in the scene (Wang et al., 2023). InstructNav further differentiates DDN from object goal navigation and from step-by-step visual-language navigation by emphasizing demand-based commonsense reasoning in unexplored environments (Long et al., 2024).

A recurring source of confusion is acronym collision. In DynaNav, a terminology note states that the paper does not explicitly use the term “Demand-Driven Navigation,” but interprets dynamic feature and layer selection based on scene complexity as implementing DDN principles of allocating computation on demand (Wang et al., 26 Sep 2025). By contrast, the ANSFL review explicitly understands DDN as Data-Driven Navigation, meaning learning-based augmentation or replacement of components in inertial navigation and sensor-fusion pipelines (Klein, 2022). The same acronym therefore spans human-demand satisfaction, on-demand conversational wayfinding, adaptive compute allocation, and learning-based navigation estimation.

2. Foundational embodied formulations

In the embodied formulation introduced on AI2-THOR with ProcThor, each episode begins with an agent pose in a mapless indoor scene and a demand instruction in natural language. The agent receives RGB observations, acts from $$A=\{\text{MoveAhead}, \text{RotateRight}, \text{RotateLeft}, \text{LookUp}, \text{LookDown}, \text{Done}\}$$, and terminates either when Done is executed or at a step cap $T=100$. The central representation is a demand-conditioned object attribute space, with object embedding $a(o)\in\mathbb{R}^k$, demand embedding $f_d(d)\in\mathbb{R}^k$, and matching score

$$m(o \mid d)=\frac{a(o)^\top f_d(d)}{\|a(o)\|\,\|f_d(d)\|}$$

or, alternatively, a learned bilinear form. Success couples navigation and selection: the chosen object must be in view and within $c_{navi}=1.5\,\text{m}$, and the selected box must satisfy $\operatorname{IoU}(b,b_{gt})\ge c_{sele}=0.5$ for some ground-truth box of an object satisfying the demand (Wang et al., 2023).

InstructNav casts DDN inside a broader generic instruction navigation system. At each step the robot observes an egocentric RGB image $V$, depth $D_t$, and pose $P_r$, and Dynamic Chain-of-Navigation converts the instruction into an updated sequence of action–landmark pairs. Control is mediated by four value maps—semantic, action, trajectory, and intuition—combined as

$T=100$0

with obstacle areas zeroed, and the next waypoint selected by

$T=100$1

This formulation is explicitly designed for demands such as “I am thirsty,” where commonsense reasoning over likely objects and functional areas is required (Long et al., 2024).

CogDDN formalizes DDN as a POMDP $T=100$2 in unknown and unstructured indoor environments using only a single front-facing camera. Observations comprise the user demand $T=100$3 and the current RGB image $T=100$4, actions are the AI2-THOR discrete navigation set plus Done, and a task succeeds once the agent is within $T=100$5 of a matched target object and issues Done or reaches proximity-triggered success. Evaluation includes SPL,

$T=100$6

together with Navigation Success Rate and Selection Success Rate (Huang et al., 15 Jul 2025).

3. Preference, multi-object, and multi-demand generalizations

Single-demand DDN assumes that one demand can be fulfilled by one acceptable object instance. MO-DDN generalizes this to multi-object search and personal preferences. Each instruction decomposes into a basic demand and a preference component, $T=100$7, and the benchmark provides sets of valid basic solutions $T=100$8 and preferred solutions $T=100$9, each defined at the category level. Satisfaction is partial rather than binary: for a candidate solution $a(o)\in\mathbb{R}^k$0 and found list $a(o)\in\mathbb{R}^k$1, $a(o)\in\mathbb{R}^k$2. The scalar episode objective is

$a(o)\in\mathbb{R}^k$3

where $a(o)\in\mathbb{R}^k$4 and $a(o)\in\mathbb{R}^k$5 are the maximum partial satisfactions over $a(o)\in\mathbb{R}^k$6 and $a(o)\in\mathbb{R}^k$7 respectively. The proposed C2FAgent uses attribute compatibility at both object and block level, with

$a(o)\in\mathbb{R}^k$8

and block score $a(o)\in\mathbb{R}^k$9 (Wang et al., 2024).

TP-MDDN extends this further to long-horizon navigation involving multiple sub-demands with explicit task preferences. BreakLLM decomposes a long instruction $f_d(d)\in\mathbb{R}^k$0 into subtasks $f_d(d)\in\mathbb{R}^k$1, each with preferences or constraints; LocateLLM selects the next target under partial knowledge; StatusMLLM monitors completion; MASMap supplies spatial memory; Dual-Tempo action generation combines slow map-based planning with fast low-level control; and an Adaptive Error Corrector handles failures in real time. A natural objective reported in the benchmark is

$f_d(d)\in\mathbb{R}^k$2

with weighted success score $f_d(d)\in\mathbb{R}^k$3. Unlike earlier field-of-view success conventions, TP-MDDN uses a distance-based success definition within $f_d(d)\in\mathbb{R}^k$4 (Li et al., 21 Nov 2025).

4. Representations, maps, and decision architectures

A central design axis in DDN is the representation of demand. The original demand-conditioned attribute-space approach uses GPT-3.5 to generate language-grounding mappings, BERT to encode the demand, CLIP text and image encoders to align textual and visual semantics, DETR to provide top-$f_d(d)\in\mathbb{R}^k$5 proposals with $f_d(d)\in\mathbb{R}^k$6, a 6-layer Transformer attribute module, a ViT-Large global image backbone, and a 2-layer LSTM policy head. The key idea is that objects satisfying the same demand share latent functional attributes even when category names differ (Wang et al., 2023).

InstructNav replaces end-to-end demand embeddings with explicit language-to-map grounding. Dynamic Chain-of-Navigation updates action–landmark pairs at every decision step, while Multi-sourced Value Maps encode semantic proximity, action priors, trajectory novelty, and multimodal intuition. Semantic maps are built from GLEE segmentations lifted into 3D, trajectory maps reward spatial novelty, and the intuition map projects a GPT-4V direction judgment onto navigable regions. The result is a zero-shot architecture that turns demand reasoning into waypoint selection without pre-built maps (Long et al., 2024).

CogDDN reframes DDN from data-driven to cognition-driven navigation. Perception is supplied by UniMODE monocular 3D detection. A supervised fine-tuned Qwen2-VL-7B performs demand matching, identifying key objects consistent with the instruction. A dual-process decision system then switches between a Heuristic Process for exploration or exploitation and an Analytic Process that reflects on hindrances, stores corrected samples in a knowledge base, and periodically distills them back into the Heuristic Process by supervised fine-tuning. Chain-of-Thought is used in all reasoning steps, and the knowledge base is seeded by A*-generated trajectories before online augmentation (Huang et al., 15 Jul 2025).

Navigation Pixie demonstrates a distinct DDN architecture for user-generated virtual worlds. Its loosely coupled design separates platform specifics from cognition through an Agent Driver and an Agent Core. The driver normalizes environmental data capture, event propagation, avatar control, and communication access; the core formats structured spatial metadata into JSON context, uses GPT-4o Realtime for speech input/output and GPT-4o for decision-making and function calling, and converts tool calls such as move_to(destination_id), say(text), and ask_clarifying(question) into driver commands. Movement is executed by Unity NavMesh Agent, while an explicit state machine cycles through suspend, waiting, player voice listening, agent thinking, agent voice playback, and action execution (Yanagawa et al., 5 Aug 2025).

A further architectural branch treats demand as a property of spatial complexity itself. DGNav addresses “Granularity Rigidity” in VLN-CE by adapting topological graph density to the angular dispersion of waypoint candidates. Candidate nodes are merged when their distance is below a threshold $f_d(d)\in\mathbb{R}^k$7, and $f_d(d)\in\mathbb{R}^k$8 is reduced in high-dispersion, high-uncertainty regions. Dynamic connectivity is introduced through

$f_d(d)\in\mathbb{R}^k$9

and this edge bias is injected into Graph Transformer attention. This is not the same task as demand-conditioned object search, but it preserves the demand-driven principle that resources and topology are modulated online rather than fixed in advance (Peng et al., 29 Jan 2026).

5. Benchmarks, datasets, and empirical findings

The scale of DDN evaluation has expanded rapidly. NavRAG generates 2,115,019 user-demand navigation instructions across 861 scenes, with mean instruction length 29.11 tokens, and pairs them with more than 10M shortest-path trajectories. On NavRAG Val Unseen, DUET trained on NavRAG reached NE 7.7 / OSR 50.0 / SR 30.7 / SPL 25.4, while HAMT trained on NavRAG reached 8.3 / 42.5 / 25.1 / 20.4; DUET pretrained on NavRAG and fine-tuned on REVERIE without object boxes reported NE 5.5 / OSR 70.7 / SR 57.3 / SPL 42.0 (Wang et al., 16 Feb 2025).

On the original embodied DDN benchmark built on AI2-THOR and ProcThor, the demand-conditioned attribute-space model reported, in the unseen scene and unseen instruction split, NSR 16.1%, NSPL 8.4%, and SSR 6.0% (Wang et al., 2023). InstructNav reported TL 4.44, SR 30.0, and SPL 14.2 under unseen scenes and unseen instructions, and states that it surpassed the previous SOTA method by 86.34% on DDN (Long et al., 2024). CogDDN, evaluated in closed loop with a single front-facing RGB camera, reported NSR 34.5, SPL 17.1, and SSR 27.5 on unseen scenes and unseen instructions, and the paper states that it outperforms single-view camera-only methods by about 15% (Huang et al., 15 Jul 2025).

For generalized formulations, MO-DDN reported that C2FAgent with the LLM branch achieved, on unseen scenes and unseen tasks, SR_b 15.88 / SR_p 9.48 and SPL_b 6.03 / SPL_p 4.50, outperforming end-to-end ObjectGoal Navigation baselines and modular LLM baselines in that benchmark (Wang et al., 2024). TP-MDDN reported, averaged over three runs on 50 tasks for each method, SR 32.00, ISR 62.89, ISPL 44.19, and STL 20.11 for AWM-Nav, compared with SR 16.00, ISR 44.67, ISPL 40.66, and STL 15.50 for DDN, and SR 16.00, ISR 42.44, ISPL 39.41, and STL 9.50 for InstructNav (Li et al., 21 Nov 2025).

In commercial virtual environments, Navigation Pixie evaluated DDN on Cluster with 99 PC client and 94 VR-HMD participants across a museum and a sky cafe. The paper reports that Group A, the on-demand DDN condition, significantly exceeded fixed-route and control conditions in dwell time and free exploration time; the summary states 1.5–1.7× longer dwell times and 3–5× longer free exploration versus fixed-route and no-agent baselines. Response time from user speech end to agent speech start was 6.40 ± 3.61 s, and VR-HMD results showed environment-dependent increases in “Anthropomorphism,” “Animacy,” and “Perceived Intelligence” in the social cafe setting (Yanagawa et al., 5 Aug 2025).

6. Extensions in transportation, adaptive computation, and data-driven navigation

Transportation work uses DDN to denote navigation or dispatch driven by spatio-temporal demand models. In on-demand transit, daily trip production and trip distribution at the Dissemination Area level are modeled as classification tasks over low, medium, and high demand levels. The best production model is Bagging with 64% test accuracy, while the best distribution model is Random Forest with 72% test accuracy. SHAP analysis identifies origin variables consistent with commercial/industrial land-use characteristics as important for production, and destination demographics as dominant for distribution; the paper explicitly recommends locating idle vehicles in the neighbourhoods with commercial/industrial land-use and using the spatio-temporal demand models obtained in this work to continuously update the operating fleet size (Alsaleh et al., 2020).

A more formal city-scale version appears in robust taxi dispatch. Demand uncertainty is represented through data-driven box or second-order cone uncertainty sets derived from hypothesis testing and bootstrap, and the dispatch objective balances idle mileage against demand-supply fairness. On four years of New York City taxi data, selecting a probabilistic guarantee level of 75% reduced average demand-supply ratio error by 31.7% and reduced average total idle driving distance by 10.13%, or about 20 million miles annually, relative to non-robust dispatch solutions (Miao et al., 2016).

Other work generalizes the “demand-driven” label from human needs to computational demand. DynaNav couples a trainable hard feature selector with an early-exit Transformer decoder, so that easy scenes use fewer pixels, tokens, and layers, while harder scenes trigger deeper computation. Relative to ViNT, DynaNav reports a 2.26× reduction in FLOPs, 42.3% lower inference time, and 32.8% lower memory usage, while improving navigation performance across four public datasets (Wang et al., 26 Sep 2025). The separate ANSFL review then uses DDN to mean Data-Driven Navigation, surveying 24 experimentally validated algorithms in AUVs, quadrotors, pedestrians, self-driving cars, and AHRS systems; reported gains include 60% AUV velocity RMSE reduction, quadrotor pure-INS distance RMSE reduced from 40–120 m to <4 m, and 25–40% improvements in adaptive fusion and AHRS settings (Klein, 2022).

7. Limitations, misconceptions, and research directions

A persistent misconception is that DDN is simply object-goal navigation with freer language. The literature instead shows that DDN introduces substitute objects, attribute or affordance matching, preference conditioning, multi-step decomposition, and, in some settings, conversational clarification. Another misconception is that DDN denotes one task family; the cited work instead spans object-seeking, conversational wayfinding, dispatch under demand uncertainty, adaptive topology, adaptive computation, and data-driven sensor fusion.

Reported limitations are correspondingly heterogeneous. Navigation Pixie identifies LLM latency and errors, speech synthesis quality, occasional spatial misrecognitions, metadata completeness, and privacy-sensitive longitudinal memory as open issues (Yanagawa et al., 5 Aug 2025). CogDDN reports short-term memory only in Explore, computationally expensive GPT-4 exploration, re-fine-tuning inefficiency when the knowledge base grows, and vulnerability to detection errors or ambiguous instructions (Huang et al., 15 Jul 2025). Early embodied DDN notes ambiguous or overly abstract demands, the dependence of latent attributes on LLM-generated co-occurrence, and RGB-only failures under occlusion or missed detections (Wang et al., 2023). MO-DDN highlights diffuse attribute vectors for overly general demands, noisy CLIP features, and the fact that the agent does not learn when to issue Done (Wang et al., 2024). TP-MDDN reports reliance on strong long-context reasoning, possible inadvertent switching between slow and fast tempos, and semantic-map false positives or negatives (Li et al., 21 Nov 2025). DGNav adds calibration-transfer issues for $$m(o \mid d)=\frac{a(o)^\top f_d(d)}{\|a(o)\|\,\|f_d(d)\|}$$0 and $$m(o \mid d)=\frac{a(o)^\top f_d(d)}{\|a(o)\|\,\|f_d(d)\|}$$1 and notes that no explicit collision penalty or reinforcement-learning safety objective is used (Peng et al., 29 Jan 2026).

The most consistent future directions are therefore not a single algorithmic recipe but a set of converging research needs: richer preference models, stronger long-horizon memory, interactive clarification, more reliable open-vocabulary perception, explicit affordance graphs, multi-agent or multi-user coordination, lighter on-device models, and better standardization of what “demand-driven” means in each subfield. The available evidence indicates that DDN is becoming a broad organizing idea for navigation systems that must adapt not only to geometry, but also to intent, uncertainty, social context, and operational constraints.