Beyond Thinking: Imagining in 360$^\circ$ for Humanoid Visual Search
Proposes 'Imagining in 360°' framework with probabilistic spatial priors, boosting humanoid visual search success by over 30%.
Key Findings
Methodology
The approach introduces an 'Imaginator' that predicts spatial priors as a probabilistic map of observed and unobserved regions, enabling single-step semantic layout inference. Multiple hypotheses are sampled via temperature sampling, providing the 'Actor' with a distribution of spatial cues. Training uses a large-scale pseudo-label pipeline, generating over 1.96 million samples by predicting entire scene layouts from limited views, bypassing costly multi-turn chain-of-thought annotations. The architecture decouples reasoning from action, allowing flexible model integration and scalable data generation.
Key Results
- On H*Bench, integrating the Imaginator with Qwen3-VL 235B increased HOS success rate from 23.62% to 65.04%, and HPS from 21.44% to 38.25%, outperforming pure Actor models by over 30%.
- Across multiple open-source and proprietary models, the spatial prior-guided framework consistently improved performance, especially in complex, real-world scenarios.
- The automatic pseudo-labeling pipeline enabled training on approximately 1.96 million samples, significantly reducing annotation costs while maintaining high accuracy.
Significance
This work addresses the core challenge of efficient spatial reasoning in embodied agents, moving beyond traditional multi-step chain reasoning. Explicit probabilistic modeling of spatial priors enhances robustness against environmental uncertainty, enabling more active and confident exploration. The scalable data pipeline democratizes large-scale training, paving the way for practical deployment of autonomous systems in complex, real-world environments. It marks a significant step forward in integrating semantic spatial understanding into embodied AI.
Technical Contribution
The key innovation is the 'Imaginator' model, which predicts scene layouts as probabilistic spatial priors, decoupling reasoning from action. This enables single-step, multi-hypothesis inference, improving robustness and interpretability. The training pipeline leverages large-scale pseudo-labeling, avoiding costly trajectory annotations. The architecture is model-agnostic, compatible with various large multimodal models, and significantly enhances search efficiency and scalability, setting new benchmarks in embodied visual reasoning.
Novelty
This is the first work to explicitly model scene layouts as probabilistic spatial priors for active search, rather than relying solely on implicit or pixel-level extrapolation. The decoupling of imagination and action, combined with scalable pseudo-label training, introduces a new paradigm for embodied AI, enabling large-scale, robust spatial reasoning in complex environments.
Limitations
- The model's performance degrades in highly dynamic or occluded environments where static spatial priors are less reliable.
- Sampling-based hypotheses, while improving robustness, may still mislead in environments with extreme uncertainty or novel layouts.
- Dependence on pseudo-labels, although scalable, may introduce biases or inaccuracies in certain scene types, affecting generalization.
Future Work
Future directions include integrating temporal dynamics for better adaptation to changing environments, enhancing multi-modal fusion for richer scene understanding, and applying reinforcement learning to optimize exploration strategies based on learned priors.
AI Executive Summary
Autonomous agents capable of efficiently exploring complex 360° environments have long been a goal in robotics and AI. Traditional methods rely heavily on multi-turn chain-of-thought reasoning, which demands extensive data annotation and incurs high computational costs. These approaches often struggle with environment uncertainty and limited visual context, leading to inefficient search behaviors. To address these challenges, this work introduces a novel 'Imagining in 360°' framework that explicitly models spatial priors as probabilistic scene layouts.
The core innovation lies in the 'Imaginator' module, which predicts the semantic structure of both observed and unobserved regions in a single inference step. Unlike pixel-level scene reconstruction, this model operates directly within a semantic language space, enabling efficient and interpretable predictions. By sampling multiple hypotheses through temperature-controlled stochasticity, the system provides the 'Actor' with a distribution of plausible spatial cues, significantly improving active search performance.
Training leverages a large-scale pseudo-label pipeline, where scene layouts are automatically generated from limited views, resulting in over 1.96 million high-quality samples. This approach circumvents the high costs associated with manual trajectory annotation, enabling scalable data generation. Experimental results on the H*Bench benchmark demonstrate that integrating the 'Imaginator' boosts success rates by over 30% across diverse models, including open-source and commercial systems. The framework's generality and efficiency mark a substantial advance in embodied visual reasoning.
Beyond performance gains, this work offers a new paradigm for spatial understanding in autonomous systems. By explicitly modeling scene layouts probabilistically, agents can explore more actively and confidently, even in cluttered or uncertain environments. The scalable training pipeline opens avenues for large-scale deployment in real-world applications such as robotics, virtual assistants, and autonomous navigation. Future work will focus on dynamic scene adaptation, multi-modal integration, and reinforcement learning to further enhance exploration capabilities. Overall, this research sets a new standard for intelligent, scalable embodied AI in complex environments.
Deep Analysis
Background
近年来，机器人自主探索逐步成为研究热点，早期依赖视觉显著性和场景上下文（如Oliva等，2003）实现目标定位。随着多模态大模型（如Wu和Xie，2024）崛起，静态图像理解取得突破，但在动态、三维空间中的主动探索仍受限。人形视觉搜索（HVS）旨模仿人类主动转头、调整视角的行为，解决遮挡和环境复杂性带来的挑战。早期工作如Yu等（2025）提出基于链式推理的连续决策模型，但面临推理成本高、数据标注难题。近年来，空间推理、场景外推、语义布局等技术不断发展，但多依赖像素级重建或隐空间，难以实现高效、可解释的空间理解。
Core Problem
核心问题在于如何在有限视野下高效、准确地推断未观察区域的空间布局。传统多轮链式推理成本高，易累积误差，且对环境变化敏感。缺乏显式空间先验模型，导致探索不主动或误导，限制自主系统在复杂场景中的应用。解决方案需引入高效空间推断机制，降低数据成本，提升鲁棒性，满足实际部署需求。
Innovation
本研究的创新在于提出“Imaginator”模型，将空间想象作为概率预测任务，显式生成场景布局。通过单步采样多重假设，为行动者提供丰富空间信息，增强决策能力。架构上，推理与行动解耦，利用自动伪标签实现大规模训练，突破长轨迹标注瓶颈。模型兼容多模态系统，验证其广泛适用性，为自主探索提供新范式。
Methodology
- �� 构建空间布局预测器（Imaginator），输入有限视野图像和指令，输出观察区域和未观察区域的语义坐标。• 采用单步无状态推理，结合温度采样生成多重空间假设，捕获环境不确定性。• 训练利用自动伪标签，基于全景场景的空间掩码，生成超过196万样本，避免长轨迹标注成本。• 设计建议转换器，将空间坐标转化为相对动作指令，指导行动者。• 行动者模型（Actor）依据空间建议，执行旋转或提交目标，完成搜索任务。• 采样策略逐步收敛，随着探索深入，环境不确定性降低，提升搜索效率。
Experiments
采用H*Bench数据集，涵盖多样场景，包括城市、室内等，评估搜索成功率。对比纯Actor模型与结合Imaginator的架构，验证性能提升。关键指标为目标定位成功率和搜索步数。模型在不同规模的预训练数据上进行训练，调优采样温度和假设数。通过消融实验验证空间布局预测的必要性和采样策略的有效性。结果显示，结合空间先验后，模型在HOS任务中成功率提升超过30%，在复杂环境中表现更稳健，验证了空间想象的有效性。
Results
结合Imaginator的模型在H*Bench上表现优异，Qwen3-VL 235B模型成功率由23.62%提升到65.04%，HPS由21.44%提升到38.25%。多模型验证显示，空间先验显著改善搜索效率，尤其在复杂环境中表现优越。自动伪标签训练使数据规模扩大至196万样本，成本大幅降低。消融分析表明，空间布局预测和多重采样策略是性能提升的关键因素。整体而言，该方法在多场景、多模型环境中均实现了优异的性能，验证了其广泛适用性和有效性。
Applications
该技术可应用于自主机器人、无人驾驶、虚拟现实等领域，提升环境理解和目标定位能力。在实际场景中，系统无需大量手工标注，依赖自动生成的空间布局，实现高效自主探索。未来，结合多模态信息和强化学习，有望实现更智能、更鲁棒的自主系统，推动智能制造、安防、导航等行业的发展。
Limitations & Outlook
当前模型在极端遮挡、动态变化环境中仍存在推理失误，空间先验假设在复杂场景下可能失效。采样策略虽提升鲁棒性，但在高不确定性场景中仍存在误导风险。训练依赖大量伪标签，虽然成本低，但在某些场景下可能引入偏差，影响泛化能力。未来需优化模型结构，增强动态适应性，减少对大规模伪标签的依赖。
Plain Language Accessible to non-experts
想象你在一个陌生的房间里找东西，你只能看到一小块区域。为了找到目标，你会根据已有的线索猜测房间的布局，比如门可能在左边，沙发在角落。每次你转头，都会根据新线索调整猜测，逐步缩小范围。这就像这个系统中的“想象者”模型，它会在脑海中模拟房间的全貌，预测未看到的区域，帮助你更快找到目标。这样，你不用盲目探索，而是根据已有信息合理推测，节省时间和精力。
ELI14 Explained like you're 14
你知道当你在一个陌生的房间里找东西时，可能会先猜猜门在哪里，沙发可能在哪个角落，然后转头看看是不是符合你的猜测。其实，你在用一种“想象”来帮自己找东西。这个研究也是这样，它让机器人学会在脑海中“想象”房间的样子，预测还没看到的地方长什么样。这样，机器人就能更聪明地转头，找到目标，比以前盲目转圈快多了。它就像你玩捉迷藏时，脑袋里会有个小地图，帮你更快找到藏起来的朋友。
Abstract
Humanoid Visual Search (HVS) requires agents to actively explore immersive 360$^\circ$ environments. While prior methods treat this as a monolithic task relying on cumulative, multi-turn Chain-of-Thought (CoT) reasoning, they impose heavy cognitive burdens and require expensive trajectory-level annotations. In this paper, we propose Imagining in 360$^\circ$, a novel framework that decouples the exploration process into a specialized Imaginator and an Actor. The Imaginator functions as a probabilistic predictor of spatial priors; instead of maintaining a cumulative reasoning chain, it infers the semantic layout of both observed and unobserved regions in a single step. By sampling multiple hypotheses within this semantic space, we provide the Actor with a distribution of effective spatial information, offering robust guidance that hedges against uncertainty during active search. This decoupled architecture significantly lowers data engineering costs by eliminating the need for full-trajectory CoT annotations, enabling the generation of over 1.96 million curated training samples. Extensive experiments demonstrate that explicitly modeling semantic spatial priors drastically improves search efficiency and success rates in complex, in-the-wild environments.