# SpatialEvo: Self-Evolving Spatial Intelligence via Deterministic Geometric Environments

> SpatialEvo introduces a self-evolving spatial reasoning framework that uses a deterministic geometric environment to provide exact, noise-free supervision, enabling a single vision-language model to co-evolve as both questioner and solver via spatial-grounded self-play.

- **Source:** [arXiv](https://arxiv.org/abs/2604.14144)
- **Published:** 2026-04-17
- **Permalink:** https://picx.dev/p/Go4fuT
- **Whiteboard:** https://picx.dev/p/Go4fuT/image

## Summary

# SpatialEvo: Self-Evolving Spatial Intelligence via Deterministic Geometric Environments

## Summary (Overview)
*   **Core Innovation:** Introduces the first self-evolving framework for 3D spatial reasoning, centered on a **Deterministic Geometric Environment (DGE)**. The DGE computes exact, zero-noise ground truth directly from 3D scene assets (point clouds, camera poses), replacing unreliable model consensus with objective physical feedback.
*   **Training Paradigm:** Employs a single Vision-Language Model (VLM) that **co-evolves** between **Questioner** and **Solver** roles via spatial-grounded self-play. The Questioner generates valid spatial questions, while the Solver answers them under DGE constraints.
*   **Adaptive Curriculum:** A task-adaptive scheduler dynamically concentrates training on the model's weakest spatial reasoning categories, enabling **endogenous curriculum emergence** without manual design.
*   **Key Results:** SpatialEvo achieves the highest average performance across nine benchmarks at both 3B and 7B model scales, showing strong gains on spatial reasoning tasks (e.g., VSI-Bench) **without degradation** on general visual understanding benchmarks.
*   **Fundamental Insight:** Leverages the unique property of 3D spatial reasoning—its answers are **deterministic consequences of underlying geometry**—to bypass the self-reinforcing error problem inherent in consensus-based self-evolution methods.

## Introduction and Theoretical Foundation
Spatial reasoning over 3D scenes is a core capability for embodied intelligence. However, continuous model improvement is bottlenecked by the high cost of manual geometric annotation for creating large-scale datasets. While the **self-evolving paradigm** (where models improve through iterative self-play) offers a promising path to overcome this, existing methods rely on **model consensus** (e.g., majority voting) to construct pseudo-labels for training. This introduces a systematic bias: training anchored to these labels risks **reinforcing the model's existing errors** rather than correcting them.

The paper identifies a distinctive property of **3D spatial reasoning** that circumvents this limitation: unlike natural language or general vision tasks, the ground truth for a spatial question is a **deterministic consequence of the underlying geometry**. Given a dense point cloud, calibrated camera poses, and a well-formed question, the correct answer can be computed **exactly and programmatically**, without any model judgment. This transforms every unannotated 3D scene into a potential source of **noise-free supervision**.

Building on this insight, the authors propose **SpatialEvo**, a framework that integrates VLM self-play with programmatic physical verification. Its core is the **Deterministic Geometric Environment (DGE)**, which serves as an exact geometric oracle.

## Methodology
The SpatialEvo framework consists of two core components: the **Deterministic Geometric Environment (DGE)** and **Spatial-Grounded Policy Co-Evolution**.

### 1. Deterministic Geometric Environment (DGE)
The DGE formalizes 3D spatial reasoning into executable atomic rules, transforming scene assets into a zero-noise interactive judge.

*   **Task Formalization:** The framework defines 16 spatial reasoning task categories, organized by observational granularity:
    *   **Multi-image scene-level (6 tasks):** e.g., Object Counting, Absolute Distance, Relative Direction.
    *   **Single-image (3 tasks):** e.g., Single-View Relative Direction, Camera-Object Distance.
    *   **Dual-image (7 tasks):** e.g., Inter-Camera Relative Position, Visibility Comparison.
*   **Geometric Validation Rule Sets:** For each task category, the DGE pre-defines a set of geometric verification rules that decouple spatial intuition into executable atomic criteria. Rules check:
    1.  **Premise Consistency:** All referenced scene entities (objects, frames) exist and are uniquely localizable.
    2.  **Inferential Solvability:** Geometric premises are unambiguously computable (e.g., sufficient point cloud density, meaningful viewpoint disparity).
    3.  **Geometric Degeneracy Filtering:** Edge cases with low training value are discarded.
*   **Automated Verification Pipeline:** A three-stage pipeline converts free-form questions into verified ground truth.
    1.  **Entity Parsing:** A lightweight LLM extracts structured entities (frame indices, object categories) from the question text.
    2.  **Legality Verification:** Extracted entities are validated against the task's rule set. Invalid questions are truncated.
    3.  **Ground-Truth Synthesis:** For valid questions, the DGE invokes its geometric toolkit (coordinate transforms, bounding-box fitting, depth projection) to compute the precise answer.

### 2. Spatial-Grounded Policy Co-Evolution
A single policy model $\pi_\theta$ is trained to co-evolve in two roles using the GRPO (Group Relative Policy Optimization) framework.

*   **Spatial Self-Play Mechanism:** The model alternates between **Questioner** and **Solver** via role-conditioned prompting.
    *   **Questioner:** Perceives the holistic 3D scene layout from multi-view images and generates a physically valid spatial question for a task sampled by the scheduler.
    *   **Solver:** Given a question (and its DGE-computed ground truth), derives a precise answer through step-by-step reasoning.
*   **Task.Adaptive Scheduler:** Drives an endogenous, adaptive curriculum. It maintains a historical effective accuracy $\bar{a}_k$ for each task category $k$. The sampling probability $p_k$ for a feasible task is weighted inversely to this accuracy:
    $$p_k = \frac{w_k}{\sum_{j \in \mathcal{T}_s^{\text{feasible}}} w_j}, \quad w_k = \max(\delta, 1 - \bar{a}_k)$$
    where $\delta$ is a minimum exploration weight. This focuses training on the model's weakest areas.
*   **Reward Design:**
    *   **Questioner Reward ($r_Q$):** Encourages generating valid, well-grounded questions.
        $$r_Q = \alpha f_{\text{fmt}} + (1-\alpha) f_{\text{valid}} \cdot f_{\text{obs}}$$
        where $f_{\text{fmt}}$ is format compliance, $f_{\text{valid}}$ is the DGE validity score, and $f_{\text{obs}}$ is a score for the quality of visual observation.
    *   **Solver Reward ($r_A$):** Rewards accurate answers or good explanations for invalid questions.
        $$
        r_A = \begin{cases}
        \alpha f_{\text{fmt}} + (1-\alpha) f_{\text{acc}}, & \text{if } Q \text{ is valid} \\
        \alpha f_{\text{fmt}} + (1-\alpha) f_{\text{explain}}, & \text{if } Q \text{ is invalid}
        \end{cases}
        $$
        where $f_{\text{acc}}$ is accuracy against DGE ground truth, and $f_{\text{explain}}$ is the quality of explaining why a question is invalid.

## Empirical Validation / Results
Experiments were conducted across nine benchmarks using Qwen2.5-VL-3B and 7B models. Training data was constructed from ~4K scenes from ScanNet, ScanNet++, and ARKitScenes.

### Main Results
**Table 1: Main results across nine benchmarks. SpatialEvo achieves the highest average score at both 3B and 7B scales.**

| **Benchmark** | **Qwen2.5-VL-3B** | | | | **Qwen2.5-VL-7B** | | | | |
| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| | **Baseline** | **SpatialLadder** | **SpaceR** | **SpatialSSRL** | **SpatialEvo** | **Baseline** | **ViLaSR** | **SpaceR** | **SpatialSSRL** | **SpatialEvo** |
| **VSI-Bench** | 28.1 | 45.7 | 36.0 | 28.0 | **39.2** | 31.1 | 45.4 | 36.8 | 33.7 | **46.1** |
| **RealWorldQA** | 63.4 | 57.1 | 61.4 | 65.4 | **66.5** | 69.5 | 57.9 | 64.7 | 69.9 | 66.7 |
| **EmbSpatial** | 55.9 | 57.6 | 55.6 | 59.8 | **61.2** | 63.6 | 47.8 | 60.3 | 69.3 | **66.0** |
| **SpatialViz** | 24.2 | 28.6 | 31.9 | 25.9 | 25.4 | 27.0 | 29.8 | 30.9 | 28.4 | **28.6** |
| **STARE** | 33.1 | 26.4 | 36.8 | 36.8 | **36.9** | 41.8 | 21.4 | 36.2 | 43.3 | **41.3** |
| **CoreCognition** | 56.8 | 58.3 | 29.1 | 57.6 | 57.4 | 59.6 | 56.4 | 56.4 | 60.2 | **60.2** |
| **ViewSpatial** | 36.2 | 43.0 | 35.9 | 38.4 | **42.3** | 36.4 | 32.3 | 35.1 | 37.5 | **43.2** |
| **V-STAR** | 74.9 | 36.7 | 75.4 | 77.0 | 75.4 | 78.5 | 35.6 | 73.8 | 79.1 | **78.0** |
| **MMStar** | 54.6 | 45.8 | 44.9 | 56.5 | 55.2 | 61.6 | 60.8 | 54.9 | 63.5 | **62.5** |
| **AVG** | 47.5 | 44.4 | 45.2 | 49.5 | **51.1** | 52.1 | 43.0 | 49.9 | 53.9 | **54.7** |

*   **Spatial Reasoning Gains:** SpatialEvo shows substantial improvements on core spatial benchmarks (VSI-Bench, EmbSpatial, ViewSpatial).
*   **General Capability Retention:** Performance on general visual understanding benchmarks (MMStar, RealWorldQA) remains competitive, indicating no degradation from spatial specialization.
*   **Scale Consistency:** Advantages hold at both 3B and 7B model scales.

### Ablation Studies
**Table 2: Ablation study on SpatialEvo (Qwen2.5-VL-7B). Removing physical grounding causes the largest performance drop.**

| **Variant** | **VSI-Bench** | **RealWorldQA** | **EmbSpatial** | **SpatialViz** | **STARE** | **CoreCognition** | **ViewSpatial** | **V-STAR** | **MMStar** | **Avg** | **Δ Avg** |
| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| **SpatialEvo (Ours)** | **46.1** | 66.7 | 66.0 | **28.6** | **41.3** | **60.2** | **43.2** | **78.0** | **62.5** | **54.7** | – |
| *Architecture Design* |
| w/o Questioner | 40.2 | **67.1** | 65.7 | 27.7 | 39.0 | 60.5 | 40.4 | 77.0 | 59.9 | 53.1 | ↓ 1.6 |
| w/o Solver | 36.6 | 70.2 | 61.8 | 26.1 | 39.8 | 58.5 | 34.5 | 75.4 | 60.5 | 51.5 | ↓ 3.2 |
| **w/o Physical Grounding** | 18.8 | 68.5 | 63.5 | 23.4 | 39.7 | 59.7 | 35.4 | 77.0 | 60.6 | 49.6 | **↓ 5.1** |
| w/o Adaptive Scheduler | 43.4 | 68.5 | **68.0** | 27.5 | 39.5 | 60.3 | 43.2 | 77.0 | 62.4 | 54.4 | ↓ 0.3 |

*   **Critical Role of DGE:** Replacing DGE ground truth with majority-vote pseudo-labels (`w/o Physical Grounding`) causes the largest average drop (**-5.1**) and a catastrophic collapse on VSI-Bench (from 46.1 to 18.8), validating the necessity of deterministic feedback.
*   **Importance of Co-Evolution:** Removing either the Questioner or Solver role leads to significant performance degradation, confirming the value of the dual-role, self-play mechanism.
*   **Scheduler Contribution:** The adaptive scheduler provides a consistent, albeit smaller, boost to performance.

### Analysis: Online Evolution vs. Static Learning
A controlled comparison shows that **online self-evolution outperforms static dataset training**, even when the static dataset is constructed from SpatialEvo's own offline generated data.

**Table 3: Comparison of online self-evolution and static learning paradigms on VSI-Bench (Qwen2.5-VL-3B). SpatialEvo's online RL outperforms all SFT baselines.**

| **Paradigm** | **Method** | **Obj. Count** | **Abs. Dist.** | **Obj. Size** | **Room Size** | **Rel. Dist.** | **Rel. Dir.** | **Route Plan** | **Appr. Order** | **Avg.** |
| :--- | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| **Qwen2.5-VL-3B (Base)** | | 33.5 | 21.1 | 17.9 | 22.6 | 32.8 | 42.7 | 29.9 | 21.0 | 28.0 |
| **RL** | w/ SpatialLadder RL | 62.8 | 29.8 | 59.2 | 27.8 | 38.9 | 44.2 | 33.5 | 24.9 | 40.1 |
| | **SpatialEvo (Online RL)** | **65.2** | **35.1** | **61.3** | **51.4** | **46.2** | **44.3** | 29.9 | **26.5** | **46.3** |
| **SFT** | w/ SpatialLadder Data | 63.0 | 31.9 | 61.2 | 43.0 | 43.0 | 43.6 | 32.5 | **31.7** | 43.7 |
| | w/ SpaceR Data | 28.3 | 25.9 | 36.3 | 34.9 | 35.1 | 46.8 | **35.1** | 43.4 | 36.3 |
| | w/ SpatialSSRL Data | 35.6 | 24.6 | 15.4 | 21.8 | 34.1 | 39.2 | 28.9 | 23.5 | 28.1 |
| | **w/ SpatialEvo Offline Data** | 62.6 | 28.9 | 60.0 | 49.1 | 40.3 | 45.7 | 31.4 | 25.9 | 43.9 |

This demonstrates the key advantage of **online self-evolution**: the training distribution dynamically aligns with the model's current cognitive frontier through real-time DGE interaction, enabling adaptive hard-sample mining that static datasets cannot replicate.

### Curriculum Emergence Analysis
Training dynamics confirm the endogenous emergence of an adaptive curriculum.
*   **Reward Trajectories:** Both Questioner and Solver rewards improve steadily. The Questioner quickly learns to generate valid questions ($f_{\text{valid}} \rightarrow 1.0$), while the Solver's accuracy reward increases.
*   **Task Sampling Shifts:** The adaptive scheduler successfully up-weights harder categories (e.g., Relative Direction from a uniform 16.7% to 21.8%) and down-weights easier ones (e.g., Room Area to 12.5%) as training progresses.

## Theoretical and Practical Implications
*   **Paradigm Shift for Self-Evolution:** SpatialEvo demonstrates that in domains where **ground truth is physically deterministic**, self-evolution can be anchored to **objective environmental feedback** rather than noisy model consensus. This prevents error reinforcement and enables truly corrective learning.
*   **Data-Efficient Continuous Improvement:** The framework provides a scalable path for continuous model improvement in 3D spatial reasoning without proportional investment in human annotation. It turns unannotated 3D scene datasets into perpetual training engines.
*   **Embodied Intelligence Foundation:** The physically grounded self-evolution paradigm can serve as a reference for broader embodied AI research, where an agent's capabilities emerge from interaction with a verifiable physical world.
*   **Limitations:** The approach currently depends on high-fidelity 3D assets (dense point clouds, calibrated poses), limiting it to static indoor environments. Performance is also sensitive to the quality of entity parsing and the underlying point cloud reconstruction.

## Conclusion
SpatialEvo introduces the

---

_Markdown view of https://picx.dev/p/Go4fuT, served by PicX — AI-generated visual whiteboard summaries of research papers._