GS-Reasoner: Unified Spatial Reasoning Framework

• GS-Reasoner is a unified framework that combines explicit geometric representations with semantic, logical, and probabilistic inference for spatial reasoning.
• It constructs detailed scene graphs using oriented bounding boxes and spatial predicates to support applications in robotics, AR/XR, and geospatial analysis.
• Experimental results demonstrate state-of-the-art performance on benchmarks for visual grounding, spatial QA, and robotic manipulation tasks.

A Grounded-Spatial Reasoner (GS-Reasoner) is a unified framework for spatial reasoning that combines explicit geometric grounding with semantic, logical, and probabilistic inference, enabling robust and interpretable spatial question answering, visual grounding, and decision-making in environments ranging from 2D maps to large-scale 3D scenes and remote sensing imagery. This paradigm encapsulates mathematical representations for objects, multi-layered scene graph abstractions, deterministic or neural-symbolic spatial computations, and tightly-coupled reasoning mechanisms, forming a backbone for tasks in robotics, AR/XR, embodied manipulation, geospatial analysis, and multimodal agent systems (Chen et al., 15 Oct 2025, Hu et al., 2 Feb 2026, Yao et al., 25 Jul 2025, Ropero et al., 16 Mar 2026, Jahangard et al., 30 Oct 2025, Sharma, 13 Apr 2026, Häsler et al., 25 Apr 2025, Janner et al., 2017, Deruyttere et al., 2020).

1. Mathematical Representations and Scene Graph Abstractions

At the core of GS-Reasoner is the rigorous representation of spatial environments. Each spatial entity is parameterized as an Oriented Bounding Box (OBB) in $\mathbb{R}^3$ or a comparable geometric primitive, defined by its center $\mathbf{c}\in\mathbb{R}^3$, orientation (rotation matrix $R\in SO(3)$ or quaternion $q$), and half-extents $\mathbf{h}=(h_x, h_y, h_z)$. Objects' occupancy sets are denoted as $$O = \{\mathbf{c} + R\ \mathrm{diag}(\mathbf{h})\mathbf{u}\ |\ \mathbf{u} \in [-1,1]^3\}$$ (Häsler et al., 25 Apr 2025, Chen et al., 15 Oct 2025).

GS-Reasoners construct scene graphs $G=(V,E)$ where:

• $V$ is the set of object nodes, each storing semantic properties (labels, attributes), geometry (OBB, volume, pose), and zone/taxonomic information.
• $E$ encodes binary or higher-order spatial predicates, such as adjacency, containment, orientation, 'on/near/behind', with relations computed as deterministic Boolean or metric-valued functions of object geometry (Häsler et al., 25 Apr 2025, Sharma, 13 Apr 2026, Jahangard et al., 30 Oct 2025, Chen et al., 15 Oct 2025).

For robotics and manipulation, node embeddings $\mathbf{v}_i \in \mathbb{R}^{d_v}$ and relation embeddings $\mathbf{c}\in\mathbb{R}^3$0 support message-passing and relational reasoning within a differentiable, structured latent space (Hu et al., 2 Feb 2026, Jahangard et al., 30 Oct 2025).

2. Spatial Knowledge Graphs and Predicate Formalism

GS-Reasoner supports an extensive formal ontology of spatial predicates ($\mathbf{c}\in\mathbb{R}^3$1), each defined via geometric, topological, or directional rules:

• On: $\mathbf{c}\in\mathbb{R}^3$2
• Inside: $\mathbf{c}\in\mathbb{R}^3$3
• Near: $\mathbf{c}\in\mathbb{R}^3$4
• Behind: $\mathbf{c}\in\mathbb{R}^3$5
• Over 80 such predicates, encompassing adjacency, connectivity, directionality, and more, are available to drive inference and logical queries (Häsler et al., 25 Apr 2025).

Predicates become edge-types in the scene graph, enabling pattern-matching, dynamic rule evaluation, and external semantic queries via compatibility with RDF/OWL triple stores.

3. Perception, Grounding, and Multi-Modal Input Integration

GS-Reasoner frameworks interface tightly with high-capacity perception modules:

• Panoramic image and point cloud fusion: Entities are detected, attributed, and their 3D centers or volumes estimated using detectors (e.g., Florence, InternVL, SigLIP ViT, Point Transformer), as in robotics and AR (Jahangard et al., 30 Oct 2025, Chen et al., 15 Oct 2025).
• Free-form or structured queries: Natural language instructions or commands are parsed and tokenized, then mapped into reasoning instructions within the semantic space of the system (Yao et al., 25 Jul 2025, Chen et al., 15 Oct 2025, Janner et al., 2017).
• All detected entities and relations are compiled into the scene graph or spatial knowledge graph, which is then exposed for symbolic or neural reasoning.

This perception-reasoning separation is essential for geometric grounding, as evidenced by RieMind and similar systems, which isolate the reasoning process from perception noise and enable high interpretability and analyzability (Ropero et al., 16 Mar 2026).

4. Reasoning Mechanisms and Computational Pipelines

GS-Reasoners implement a variety of inference and decision workflows:

• Pipeline-style deterministic reasoning: Linear or Datalog-like pipelines of filters, deduce, produce, and map steps operate on the spatial knowledge base. Each stage performs transformations or pattern matches using spatial predicates (e.g., filter(type=="chair") | pick(near AND infront) | produce(on:label="seat-zone")) (Häsler et al., 25 Apr 2025).
• Neural-symbolic reasoning: Hybrid systems couple neural feature encoding with explicit logical rule evaluation, using scene graphs as the substrate for graph traversal, attribute filters, and relational validation (Jahangard et al., 30 Oct 2025, Chen et al., 15 Oct 2025).
• Agentic tool-based protocols: Agent LLMs interact with deterministic geometric service APIs (e.g., geom_distance, loc_project), issuing structured queries and consuming structured outputs per reasoning step, ensuring that spatial reasoning is grounded in explicitly computed geometric facts (Ropero et al., 16 Mar 2026).
• Transition-based reasoning for manipulation and planning: Scene graphs are updated over time via action-conditioned graph neural networks and symbolic effect functions, encoding conditions, feasibility, and goal satisfaction (Hu et al., 2 Feb 2026).

A crucial methodological advance is the offloading of all deterministic sub-problems (distance, adjacency, containment, collision) to geometric engines, restricting LLM inference to the residual, under-constrained aspects of the task. This architecture ensures non-hallucinated spatial reasoning (Sharma, 13 Apr 2026).

5. Learning Algorithms, Supervision, and Reward Schemes

GS-Reasoners are trained by combinations of:

• Supervised fine-tuning (SFT): Systems are optimized via token-level or structured prediction losses, reconstructing ground-truth scene graphs, action sequences, or reasoning traces (Chen et al., 15 Oct 2025, Hu et al., 2 Feb 2026, Jahangard et al., 30 Oct 2025).
• Reinforcement learning (RL): Policies are induced via policy-gradient objectives (e.g., GRPO) leveraging spatially dense rewards such as IoU, continuous positional reward, and format compliance, as in remote sensing grounding and visual reasoning (Yao et al., 25 Jul 2025, Huang et al., 29 Jan 2026).
• Spatial consistency guidance: Group-based rollout statistics regularize the policy (e.g., variance adaptive weights to stabilize spatial predictions across sampled outputs) (Huang et al., 29 Jan 2026).
• Synthetic chain-of-thought (CoT) data: Large-scale datasets with interleaved bounding box annotations and CoT reasoning are key for priming end-to-end models to link explicit grounding with subsequent semantic inference (Chen et al., 15 Oct 2025, Huang et al., 29 Jan 2026).

Hierarchical learning objectives combine scene grounding, world-modeling, action planning, and goal fulfillment losses, ensuring multi-task spatial generalization (Hu et al., 2 Feb 2026).

6. Experimental Results and Application Domains

GS-Reasoner variants demonstrate leading empirical performance across a spectrum of benchmarks:

Task / Benchmark	GS-Reasoner Performance	Baselines	Key Metric(s)
3D Visual Grounding (ScanRefer)	Acc@25=60.8%, Acc@50=42.2%	Mesh-prop VLMs	Accuracy@threshold
Spatial QA (VSI-Bench, pred depth)	64.7%	VLM-3R: 60.9%	Avg. correct response
Region-level (EarthReason, remote)	[email protected]=68.1%, gIoU=69.3	Qwen2.5-VL-7B: 45.8%	[email protected], gIoU
Zero-shot manipulation (RLBench)	92.5% task progress	GPT-5: 80.0%	Task progress (TP %)
JRDB-Reasoning (Robotics, mAP)	35.7 (fine-grained)	Best: 12.18	mAP@[.50:.95], mIoU
Remote Sensing Visual Grounding	[email protected]=71.8% (DIOR-RSVG)	Baseline: 66.7%	[email protected], mIoU

GS-Reasoner systems are applied in autonomous manipulation, AR/XR spatial augmentation, real-time robotics navigation, remote sensing query answering, and as spatial reasoning cores for agentic LLMs and self-driving frameworks (Jahangard et al., 30 Oct 2025, Yao et al., 25 Jul 2025, Sharma, 13 Apr 2026, Huang et al., 29 Jan 2026).

7. Interpretability, Limitations, and Outlook

Explicit geometric grounding enables transparent, auditable reasoning traces: each sub-answer is rooted in computed facts (bounding boxes, predicates, collision detections). The use of chain-of-thought with explicit spatial references supports qualitative analysis and debugging and provides reliable proxies for model decision-making (Chen et al., 15 Oct 2025, Jahangard et al., 30 Oct 2025, Sharma, 13 Apr 2026).

Limitations include:

• Continued dependence on perception quality; noisy detection or inaccurate geometry degrades reasoning.
• For end-to-end joint models, there is a trade-off between compact unified tokenization and fine-grained spatial fidelity, with explicit geometric modules often outperforming monolithic LLM-only approaches on intricate spatial reasoning (Chen et al., 15 Oct 2025).
• Synthetic CoT datasets may not capture multi-object or relational complexities outside data distributions (Huang et al., 29 Jan 2026).

Ongoing research directions are focused on integrating dense segmentation, multi-view fusion, robust SLAM coupling, dynamic scene updates, and tighter coupling with action-planning in embodied settings (Chen et al., 15 Oct 2025, Hu et al., 2 Feb 2026).

---

GS-Reasoner methodologies anchor spatial reasoning in formally defined geometric and logical substrates, combining deterministic computation and neural-symbolic reasoning to achieve high interpretability and state-of-the-art empirical performance across spatial QA, grounding, robotics, and planetary-scale Earth observation (Huang et al., 29 Jan 2026, Yao et al., 25 Jul 2025, Chen et al., 15 Oct 2025, Sharma, 13 Apr 2026, Ropero et al., 16 Mar 2026, Jahangard et al., 30 Oct 2025, Hu et al., 2 Feb 2026, Häsler et al., 25 Apr 2025, Janner et al., 2017, Deruyttere et al., 2020).