Prompt-Guided SpatialBot System

• Prompt-Guided SpatialBot is a framework that leverages multimodal prompts (textual, visual, coordinate) to shape spatial reasoning and control across diverse tasks.
• It integrates backbone vision-language models with a prompt-construction layer that explicitly grounds spatial tasks like video QA, navigation, and object search.
• Empirical results demonstrate significant performance gains in depth reasoning, warehouse navigation, and egocentric video QA through tailored prompt conditioning.

to=arxiv_search.search 尚度=medium 北京赛车前json true content='{"^^^^^^^^13query13 OR ^^^^^^^^13ti:\13 OR ^^^^^^^^13ti:\13 OR ^^^^^^^^13ti:\13 spatial reasoning vision-language-action13^^^^^^^^", "^^^^^^^^13max_results13 ^^^^13ti:\13query13^^^^, "^^^^^^^^13sort_by13 "^^^^^^^^13submittedDate13 to=arxiv_search.search 不中返json true content='{"^^^^^^^^13query13 Dynamic Prompt Routing for Zero-Shot Spatial Reasoning\" OR ^^^^^^^^13ti:\13query13 Precise Spatial Understanding with Vision LLMs\" OR ^^^^^^^^13ti:\13query13 Keyframe-driven Zero-Shot Spatial Reasoning with Off-the-Shelf Multimodal LLMs\"", "^^^^^^^^13max_results13 ^^^^13ti:\13query13^^^^, "^^^^^^^^13sort_by13 "relevance"}' to=arxiv_search.search to=arxiv_search.search The term “Prompt-Guided SpatialBot” is used in multiple recent system descriptions to denote prompt-conditioned spatial reasoning or control pipelines in which prompts serve as operational interfaces for grounding, routing, planning, or action selection. In these systems, prompts do not merely specify tasks; they encode situational descriptions, question typologies, camera poses, semantic maps, bounding boxes, region identifiers, visual overlays, or first-frame spatial anchors. The resulting design space spans zero-shot egocentric video question answering, RGB-D spatial understanding, keyframe-and-pose prompting, warehouse relation reasoning, embodied navigation, object search, and visuomotor manipulation (&&&^^^^13query13^^^^&&&, &&&^^^^13ti:\13^^^^&&&, &&&^^^^13 OR ti:\13^^^^&&&, &&&^^^^13 OR ti:\13^^^^&&&, &&&^^^^13 OR ti:\13^^^^&&&, &&&^^^^13 spatial reasoning vision-language-action13^^^^&&&).

^^^^13ti:\13^^^^. Scope and defining characteristics

Across the cited literature, a Prompt-Guided SpatialBot is not a single model family but a recurring systems pattern: a backbone VLM, MLLM, LLM, or VLA policy is paired with a prompt-construction layer that explicitly injects spatial structure before or during inference. The prompt can be textual, visual, coordinate-based, map-based, or multimodal, but its role is consistent: to shape spatial interpretation without relying exclusively on backbone-internal representations.

Setting	Prompt signal	Representative source
Egocentric video spatial QA	Question-aware routed templates and situational grounding	SpatioRoute (&&&^^^^13query13^^^^&&&)
RGB-D spatial VLMs	RGB+depth inputs, depth-aware instructions, and optional Depth(point) calls	SpatialBot (&&&^^^^13ti:\13^^^^&&&)
Zero-shot ^^^^13 OR ti:\13^^^^D QA from image sequences	Keyframes plus serialized camera poses	SpatialPrompting (&&&^^^^13 OR ti:\13^^^^&&&)
Warehouse relation reasoning	Region IDs and bounding box coordinates embedded in text prompts	RGB-D transformer framework (&&&^^^^13query13^^^^&&&)
Visuomotor control	Crosshairs and bounding boxes rendered as visual prompts	VP-VLA (&&&^^^^13 OR ti:\13^^^^&&&)
Navigation and search	Priori-maps, linguistic maps, or LLM-derived container probabilities	PM-Nav / MapGPT / object-search planning (&&&^^^^13 OR ti:\13^^^^&&&, &&&^^^^13ti:\13 OR ti:\13^^^^&&&, &&&^^^^13 spatial reasoning vision-language-action13^^^^&&&)

A common misconception is that prompt-guided spatial systems are purely linguistic. The surveyed implementations are often explicitly multimodal: RGB-D pairs in SpatialBot, visual prompt images in VP-VLA, first-frame point or box prompts in SPOT, and map or pose serialization in PM-Nav and SpatialPrompting (&&&^^^^13ti:\13^^^^&&&, &&&^^^^13 OR ti:\13^^^^&&&, &&&^^^^13ti:\13max_results13^^^^&&&, &&&^^^^13 OR ti:\13^^^^&&&, &&&^^^^13 OR ti:\13^^^^&&&).

^^^^13 OR ti:\13^^^^. Core architectural pattern

The most direct formalization appears in SpatioRoute, where zero-shot spatial question answering over egocentric video is written as

PRESERVED_PLACEHOLDER_^^^^13query13^^^^

with routing performed before video reasoning begins. In the rule-based mode, the leading interrogative determines the template; in the LLM-driven mode, a text-only router synthesizes a task-specific prompt from PRESERVED_PLACEHOLDER_^^^^13ti:\13^^^^ with PRESERVED_PLACEHOLDER_^^^^13 OR ti:\13^^^^ few-shot demonstrations, while remaining video-free at routing time (&&&^^^^13query13^^^^&&&).

This “prompt-before-reasoning” pattern reappears in other domains. VP-VLA inserts a System ^^^^13 OR ti:\13^^^^ Planner ahead of a System ^^^^13ti:\13^^^^ Controller and renders structured visual prompts onto a separate prompt image, with the controller defined as

PRESERVED_PLACEHOLDER_^^^^13 OR ti:\13^^^^

Planner invocation is event-driven through

PRESERVED_PLACEHOLDER_^^^^13 OR ti:\13^^^^

so prompt regeneration is tied to semantic phase changes such as gripper-state transitions rather than to every frame (&&&^^^^13 OR ti:\13^^^^&&&).

A second recurring motif is prompt-conditioned abstraction. PM-Nav transforms functional-building maps into semantic priori-maps and then serializes topology, landmark constraints, and segment ordering into hierarchical prompts for global planning, mid-level wayfinding, and local control (&&&^^^^13 OR ti:\13^^^^&&&). MapGPT similarly builds an online linguistic topological map containing trajectory, connectivity, and place identifiers, then includes that map verbatim in each prompt to support multi-step planning and backtracking (&&&^^^^13ti:\13 OR ti:\13^^^^&&&). Object-search planning in partially known environments uses prompts differently: the LLM estimates marginal container probabilities, which are then injected into a model-based planner through Bellman-style expected-cost recursion rather than direct action generation (&&&^^^^13 spatial reasoning vision-language-action13^^^^&&&).

This suggests that Prompt-Guided SpatialBot systems are best understood as interface architectures. The prompt layer functions as a structured bottleneck between raw perception and downstream reasoning or control, often replacing the need for task-specific fine-tuning, explicit ^^^^13 OR ti:\13^^^^D sensors, or monolithic end-to-end policies.

^^^^13 OR ti:\13^^^^. Prompt representations and grounding mechanisms

Textual routing remains the simplest mechanism. SpatioRoute-R deterministically maps question typologies to specialized templates: What → details_scene, Is → step_by_step, How → details_scene, Can → scene_understanding, Which → details_scene, and Others → instruction_focused. Situational grounding is appended when egocentric orientation matters through the string: “Consider your current position and orientation in the scene based on. Situation: {situation}” (&&&^^^^13query13^^^^&&&).

Structured textual grounding can also be explicitly geometric. In the warehouse framework, masks are converted to bounding boxes and injected as strings such as “Region ^^^^13query13^^^^ within bounding box (x^^^^13ti:\13^^^^, y^^^^13ti:\13^^^^, x^^^^13 OR ti:\13^^^^, y^^^^13 OR ti:\13^^^^),” prefixed by “Given all bounding box sizes are in the form x^^^^13ti:\13^^^^y^^^^13ti:\13 OR ti:\13^^^^y^^^^13 OR ti:\13^^^^.” Training answers append the normalization suffix “In short the normalized answer is [label].” The reported ablation from SpatialBot_v^^^^13ti:\13^^^^ to SpatialBot_v^^^^13 OR ti:\13^^^^ isolates this prompt enrichment: S^^^^13ti:\13^^^^ = ^^^^13 OR ti:\13sort_by13^^^^.^^^^13max_results13query13^^^^ → ^^^^13sort_by13 OR ti:\13^^^^.^^^^13query13max_results13^^^^, a +^^^^13 OR ti:\13 spatial reasoning vision-language-action13^^^^.^^^^13 OR ti:\13sort_by13^^^^ point gain (&&&^^^^13query13^^^^&&&).

Visual prompting uses directly rendered spatial anchors rather than text-only coordinates. VP-VLA employs a crosshair PRESERVED_PLACEHOLDER_^^^^13 spatial reasoning vision-language-action13^^^^ for target objects and a bounding box PRESERVED_PLACEHOLDER_^^^^13max_results13^^^^ for goal regions, both overlaid onto an overhead camera view to form a dedicated prompt image (&&&^^^^13 OR ti:\13^^^^&&&). SPOT uses first-frame points or bounding boxes for manipulated object and target specification, encoding coordinates with Fourier features and simultaneously rendering them visually on the first frame before cross-attending them with frozen DINOv^^^^13 OR ti:\13^^^^ features (&&&^^^^13ti:\13max_results13^^^^&&&). Spatial Policy encodes subplans as tuples such as PRESERVED_PLACEHOLDER_^^^^13sort_by13^^^^, where action type, direction vector, and travel distance form a prompt-like spatial plan table that conditions both video generation and action prediction (&&&^^^^13 OR ti:\13submittedDate13^^^^&&&).

Depth and pose provide another grounding regime. SpatialBot adds depth as a co-equal modality, with millimeter-scale encoding and a Depth(point) API that can return precise metric depth when the model emits an explicit function call (&&&^^^^13ti:\13^^^^&&&). SpatialPrompting selects diverse keyframes and serializes camera positions and Euler angles alongside each image block; its prompt therefore makes viewpoint geometry a first-class textual object rather than an implicit latent variable (&&&^^^^13 OR ti:\13^^^^&&&). PM-Nav and MapGPT do the same at map scale by serializing segment connectivity and visited-place trajectories into language (&&&^^^^13 OR ti:\13^^^^&&&, &&&^^^^13ti:\13 OR ti:\13^^^^&&&).

A plausible implication is that prompt-guided spatial systems differ less by backbone family than by the spatial representation chosen for prompting: typed templates, region coordinates, depth values, visual overlays, semantic maps, or pose traces.

^^^^13 OR ti:\13^^^^. Major application domains

In egocentric video spatial question answering, SpatioRoute targets SQA^^^^13 OR ti:\13^^^^D under strict zero-shot, video-only conditions. The benchmark is built on ^^^^13max_results13 spatial reasoning vision-language-action13query13^^^^ ScanNet scenes, ^^^^13max_results13^^^^.^^^^13submittedDate13 situational descriptions, and ^^^^13 OR ti:\13 OR ti:\13^^^^.^^^^13 OR ti:\13^^^^K questions, with an official test set of ^^^^13 OR ti:\13^^^^,^^^^13 spatial reasoning vision-language-action13ti:\13query13^^^^ questions. Under exact-match evaluation and temperature ^^^^13query13^^^^.^^^^13 OR ti:\13^^^^, SpatioRoute-R reaches ^^^^13 spatial reasoning vision-language-action13query13^^^^.^^^^13 OR ti:\13^^^^% on Qwen^^^^13 OR ti:\13^^^^-^^^^13 OR ti:\13^^^^B, compared with a ^^^^13 OR ti:\13submittedDate13^^^^.^^^^13 OR ti:\13^^^^ baseline and ^^^^13 OR ti:\13sort_by13^^^^.^^^^13sort_by13^^^^ CoT condition, and is positioned as surpassing fine-tuned ^^^^13 OR ti:\13^^^^D-LLM (^^^^13 OR ti:\13sort_by13^^^^.^^^^13 OR ti:\13^^^^%), approaching fine-tuned LEO (^^^^13 spatial reasoning vision-language-action13query13^^^^.^^^^13query13^^^^%), and outperforming SpatialPrompting + Gemini-^^^^13 OR ti:\13^^^^.^^^^13query13^^^^ (^^^^13 OR ti:\13submittedDate13^^^^.^^^^13max_results13^^^^%) under strictly zero-shot video-only conditions (&&&^^^^13query13^^^^&&&).

In RGB-D spatial understanding, SpatialBot augments a standard VLM architecture with paired RGB and depth images and trains on SpatialQA and SpatialQA-E. SpatialBench reports that RGB-D SpatialBot variants achieve >^^^^13query13query13^^^^% on the “Depth” category across several backbones, while also improving positional, reaching, and counting performance over Bunny baselines; the paper further reports gains on general VLM benchmarks such as MMBench and MME cognition (&&&^^^^13ti:\13^^^^&&&). In warehouse spatial reasoning, the RGB-D transformer system fine-tuned on the Physical AI Spatial Intelligence Warehouse dataset reports a final score of ^^^^13sort_by13 OR ti:\13^^^^.^^^^13query13max_results13query13max_results13^^^^, placing ^^^^13 OR ti:\13^^^^th overall on the public leaderboard (&&&^^^^13query13^^^^&&&).

In zero-shot ^^^^13 OR ti:\13^^^^D QA without ^^^^13 OR ti:\13^^^^D-specific fine-tuning, SpatialPrompting uses keyframe selection and camera-pose serialization. On ScanQA validation with GPT-^^^^13 OR ti:\13^^^^o, it reports EM@^^^^13ti:\13^^^^ ^^^^13 OR ti:\13sort_by13^^^^.^^^^13 OR ti:\13 OR ti:\13^^^^, ROUGE-L ^^^^13 OR ti:\13 OR ti:\13^^^^.^^^^13 OR ti:\13query13^^^^, CIDEr ^^^^13submittedDate13sort_by13^^^^.^^^^13max_results13query13, and SPICE ^^^^13 OR ti:\13query13^^^^.^^^^13 OR ti:\13query13^^^^; on SQA^^^^13 OR ti:\13^^^^D test it reports ^^^^13 spatial reasoning vision-language-action13 OR ti:\13^^^^.^^^^13sort_by13 OR ti:\13^^^^ average accuracy, with category values including What ^^^^13 OR ti:\13submittedDate13^^^^.^^^^13max_results13 spatial reasoning vision-language-action13^^^^, Is ^^^^13max_results13 OR ti:\13^^^^.^^^^13 OR ti:\13max_results13^^^^, and How ^^^^13 spatial reasoning vision-language-action13 OR ti:\13^^^^.^^^^13 spatial reasoning vision-language-action13 spatial reasoning vision-language-action13^^^^ (&&&^^^^13 OR ti:\13^^^^&&&).

In navigation and object search, prompt-guided spatial systems bifurcate into map-conditioned planning and probability-conditioned search. PM-Nav reports large gains in functional buildings, including simulation SR/SPL values of ^^^^13submittedDate13submittedDate13^^^^/^^^^13sort_by13sort_by13 OR ti:\13query13^^^^ on easy tasks and real-world SR/SPL values of ^^^^13sort_by13 spatial reasoning vision-language-action13^^^^/^^^^13 spatial reasoning vision-language-action13query13^^^^.^^^^13query13ti:\13^^^^ on easy tasks, while baselines remain far lower (&&&^^^^13 OR ti:\13^^^^&&&). MapGPT reports zero-shot SR improvements of about ^^^^13ti:\13query13^^^^% on R^^^^13 OR ti:\13^^^^R and ^^^^13ti:\13 OR ti:\13^^^^% on REVERIE through linguistic maps and adaptive planning, with ^^^^13 OR ti:\13sort_by13^^^^.^^^^13sort_by13^^^^ SR on the ^^^^13sort_by13 OR ti:\13^^^^-scene R^^^^13 OR ti:\13^^^^R subset for GPT-^^^^13 OR ti:\13^^^^V (&&&^^^^13ti:\13 OR ti:\13^^^^&&&). In partially known object search, LLM-informed model-based planning improves over the fully LLM-driven baseline by as much as ^^^^13ti:\13ti:\13^^^^.^^^^13submittedDate13 and over an optimistic strategy by as much as ^^^^13 OR ti:\13query13^^^^.^^^^13 OR ti:\13^^^^%, while replay-based prompt/LLM selection reduces average cost by ^^^^13max_results13^^^^.^^^^13 spatial reasoning vision-language-action13^^^^% and cumulative regret by ^^^^13 OR ti:\13 OR ti:\13^^^^.^^^^13submittedDate13^^^^% (&&&^^^^13 spatial reasoning vision-language-action13^^^^&&&).

In manipulation and VLA control, SPOT formalizes Spatially Prompted Visual Trajectory Prediction and reports that BBox + visual prompting improves FDE ^^^^13query13^^^^.^^^^13ti:\13sort_by13 OR ti:\13query13^^^^ → ^^^^13query13^^^^.^^^^13ti:\13ti:\13 OR ti:\13sort_by13^^^^ and Pos. L^^^^13 OR ti:\13^^^^ ^^^^13query13^^^^.^^^^13query13query13ti:\13 OR ti:\13^^^^ → ^^^^13query13^^^^.^^^^13query13max_results13query13query13 relative to the no-prompt baseline (&&&^^^^13ti:\13max_results13^^^^&&&). VP-VLA reports ^^^^13 spatial reasoning vision-language-action13 OR ti:\13^^^^.^^^^13submittedDate13^^^^% average success on Robocasa-GR^^^^13ti:\13^^^^-Tabletop versus ^^^^13 OR ti:\13submittedDate13^^^^.^^^^13submittedDate13^^^^% for the QwenOFT baseline, and ^^^^13 spatial reasoning vision-language-action13submittedDate13^^^^.^^^^13 OR ti:\13^^^^% average on SimplerEnv versus ^^^^13 spatial reasoning vision-language-action13query13^^^^.^^^^13query13^^^^% for QwenOFT (&&&^^^^13 OR ti:\13^^^^&&&). Spatial Policy reports ^^^^13submittedDate13max_results13^^^^.^^^^13sort_by13 average success across ^^^^13ti:\13ti:\13^^^^ Meta-World tasks, with a ^^^^13 OR ti:\13 OR ti:\13^^^^.^^^^13query13^^^^% average improvement over the best baseline when dual-stage replanning is used (&&&^^^^13 OR ti:\13submittedDate13^^^^&&&). Prompt^^^^13 OR ti:\13^^^^Auto addresses one-shot motion prompting and reports geometry-robust automated control from a single demonstration via geometry-invariant Gaussian-process prediction (&&&^^^^13 OR ti:\13 OR ti:\13^^^^&&&).

^^^^13 spatial reasoning vision-language-action13^^^^. Empirical findings and recurring design trade-offs

One recurrent finding is that question-aware or task-aware prompting often outperforms uniform prompting. SpatioRoute explicitly reports that Chain-of-Thought via Think it Twice “consistently degrades” performance on Qwen-series models, with category-specific drops such as Can −^^^^13 OR ti:\13query13^^^^.^^^^13ti:\13 OR ti:\13^^^^% and How −^^^^13submittedDate13^^^^.^^^^13ti:\13query13 on Qwen^^^^13 OR ti:\13^^^^-^^^^13 OR ti:\13^^^^B, while Llama^^^^13 OR ti:\13^^^^.^^^^13 OR ti:\13^^^^-^^^^13ti:\13ti:\13^^^^B is an exception and improves from ^^^^13 OR ti:\13 OR ti:\13^^^^.^^^^13 OR ti:\13^^^^ → ^^^^13 OR ti:\13submittedDate13^^^^.^^^^13 OR ti:\13^^^^ under CoT (&&&^^^^13query13^^^^&&&). The common claim that more reasoning tokens always improve spatial performance is therefore not supported across architectures.

A second finding is that explicit spatial grounding in the prompt can substitute for heavier architectural changes. In warehouse reasoning, inserting box coordinates into text produces the already noted +^^^^13 OR ti:\13 spatial reasoning vision-language-action13^^^^.^^^^13 OR ti:\13sort_by13^^^^ S^^^^13ti:\13^^^^ gain (&&&^^^^13query13^^^^&&&). In VP-VLA, removing the grounding objective drops average performance from ^^^^13 spatial reasoning vision-language-action13 OR ti:\13^^^^.^^^^13submittedDate13^^^^% to ^^^^13 OR ti:\13query13^^^^.^^^^13 OR ti:\13^^^^%, replacing the crosshair with a point drops it to ^^^^13 OR ti:\13sort_by13^^^^.^^^^13 OR ti:\13^^^^%, and rendering prompts directly on the primary RGB image yields ^^^^13 spatial reasoning vision-language-action13query13^^^^.^^^^13submittedDate13^^^^% rather than ^^^^13 spatial reasoning vision-language-action13 OR ti:\13^^^^.^^^^13submittedDate13^^^^%, indicating that prompt form and injection pathway both matter (&&&^^^^13 OR ti:\13^^^^&&&).

A third finding is that prompt quality and prompt quantity interact nontrivially. SpatialPrompting reports its best ScanQA numbers with ^^^^13 OR ti:\13query13^^^^ images; reducing to ^^^^13ti:\13 spatial reasoning vision-language-action13^^^^ or ^^^^13 spatial reasoning vision-language-action13^^^^ images degrades performance, and removing annotation causes a larger collapse than removing keyframe extraction (&&&^^^^13 OR ti:\13^^^^&&&). Prompt-guided prefiltering for VLM image compression shows a different trade-off: by identifying prompt-relevant image regions with TinyCLIP and smoothing the rest, the method achieves ^^^^13 OR ti:\13 spatial reasoning vision-language-action13^^^^.^^^^13submittedDate13submittedDate13^^^^% to ^^^^13 OR ti:\13submittedDate13^^^^.^^^^13sort_by13submittedDate13^^^^% bitrate reduction while maintaining the same task accuracy across VQA benchmarks, indicating that prompts can shape not only reasoning but also upstream sensing and transmission (&&&^^^^13 OR ti:\13submittedDate13^^^^&&&).

A fourth finding is that prompt-conditioned systems remain architecture-dependent and domain-dependent. SpatialBot’s RGB-D instruction tuning improves spatial understanding and also lifts several general VLM benchmarks, but still relies on depth quality and prompt discipline (&&&^^^^13ti:\13^^^^&&&). PM-Nav’s hierarchical prompts and priori-maps are effective in functional buildings with repeated visual cues, whereas Prompt^^^^13 OR ti:\13^^^^Auto derives its strength from geometry-invariant canonicalization rather than from explicit scene perception (&&&^^^^13 OR ti:\13^^^^&&&, &&&^^^^13 OR ti:\13 OR ti:\13^^^^&&&). This suggests that “Prompt-Guided SpatialBot” is not a single algorithmic recipe but a family of prompt interfaces tailored to distinct spatial bottlenecks.

^^^^13max_results13^^^^. Limitations, misconceptions, and future directions

The literature consistently reports that prompt-guided spatial systems remain limited by their grounding substrate. SpatioRoute acknowledges a performance ceiling relative to fine-tuned ^^^^13 OR ti:\13^^^^D methods, over-specification in SpatioRoute-L, and architecture-specific prompting behavior (&&&^^^^13query13^^^^&&&). SpatialPrompting depends on pose estimation and carefully selected viewpoints; its SQA^^^^13 OR ti:\13^^^^D category profile indicates continued difficulty with direction-dependent questions tied to situated orientation (&&&^^^^13 OR ti:\13^^^^&&&). SpatialBot’s depth reasoning depends on depth quality, robust RGB-depth alignment, and prompt phrasing that emphasizes real-world rather than image-plane orientation (&&&^^^^13ti:\13^^^^&&&).

In robotics, the main bottlenecks shift toward calibration, temporal drift, and action verification. VP-VLA depends on reliable text-conditioned segmentation and prompt-image alignment (&&&^^^^13 OR ti:\13^^^^&&&). SPOT assumes accurate first-frame grounding and reports accumulation of drift over stitched trajectory chunks (&&&^^^^13ti:\13max_results13^^^^&&&). Spatial Policy identifies assembly as a particularly hard compositional task even after replanning (&&&^^^^13 OR ti:\13submittedDate13^^^^&&&). PM-Nav notes brittleness from rule-based routing and the absence of explicit ^^^^13 OR ti:\13^^^^D geometry (&&&^^^^13 OR ti:\13^^^^&&&). The ^^^^13 OR ti:\13^^^^D-grounded robotic planning framework reports that removing ^^^^13 OR ti:\13^^^^D prompt synthesis collapses TSR to ^^^^13query13^^^^% in its ablation, while removing SLM supervision causes a large TSR drop, underscoring the fragility of prompt-guided pipelines when either grounding or validation is removed (&&&^^^^13 spatial reasoning vision-language-action13query13^^^^&&&).

Another misconception is that prompt-guided spatial systems are inherently “lightweight.” Some are; SpatioRoute’s routing is negligible and video-free at routing time (&&&^^^^13query13^^^^&&&). Others are not: the warehouse RGB-D transformer reports training on ^^^^13 OR ti:\13^^^^× NVIDIA A^^^^13 OR ti:\13query13^^^^ (^^^^13 OR ti:\13submittedDate13^^^^ GB each) and about ^^^^13ti:\13 OR ti:\13sort_by13^^^^ hours per epoch under its sampled training regime, and object-search prompt selection requires multi-trial replay to identify the best prompt/LLM pair (&&&^^^^13query13^^^^&&&, &&&^^^^13 spatial reasoning vision-language-action13^^^^&&&).

The future directions named across the literature are relatively consistent. They include hybrid routing that balances specificity and generality, broader cross-architecture prompt evaluation, learned box or spatial tokens in addition to text-only coordinates, closed-loop replanning and online re-grounding, uncertainty-aware search and safety constraints, and extension of prompt-conditioned reasoning to broader situated benchmarks and real-world deployments (&&&^^^^13query13^^^^&&&, &&&^^^^13query13^^^^&&&, &&&^^^^13ti:\13max_results13^^^^&&&, &&&^^^^13 spatial reasoning vision-language-action13^^^^&&&). A plausible implication is that the long-term significance of Prompt-Guided SpatialBot lies not in replacing geometry or control, but in making spatial abstractions explicit enough that off-the-shelf foundation models can participate reliably in spatial reasoning and action.