Information Coordination as a Bridge: A Neuro-Symbolic Architecture for Reliable Autonomous Driving Scene Understanding
Proposes InfoCoordiBridge, a neuro-symbolic framework integrating multi-sensor coordination and verifiable reasoning, boosting autonomous scene understanding reliability.
Key Findings
Methodology
This paper introduces a BEV-centric neuro-symbolic architecture comprising a multi-agent perception layer, ICA module, and SSRE engine. The perception layer outputs structured facts and modality summaries from sensors like cameras, LiDAR, and radar. ICA performs coordinate normalization, entity alignment, and conflict-aware attribute fusion to produce a unified SceneSummary with provenance and uncertainty. SSRE employs a chain-of-prompt verification workflow, grounded in entity-based reasoning, to ensure factual consistency and reduce hallucinations. Specific algorithms include covariance intersection for multi-source fusion, confidence-weighted attribute fusion, and consistency gates for conflict detection. Experiments on nuScenes and Waymo datasets show the system maintains ~70% detection AP, improves fusion consistency to 98%, reduces redundancy below 1%, and enhances QA factual grounding.
Key Results
- ICA preserves ~70% 3D detection AP while boosting fusion consistency to 98%, with redundancy below 1%, outperforming BEVFusion and TransFusion.
- In QA tasks, SSRE reduces hallucinated entity mentions by 20%, improves factual accuracy, surpassing baseline VLMs and perception-only models.
- Overall, the system effectively prevents conflicting perception evidence from propagating into high-level reasoning, ensuring more trustworthy scene understanding.
Significance
This work addresses a critical gap in autonomous driving: ensuring high-level scene understanding is both semantically consistent and verifiable. By explicitly coordinating multi-sensor data and integrating a validation-driven reasoning engine, it significantly mitigates hallucination risks and enhances safety. The approach advances the state-of-the-art in trustworthy perception, providing a foundation for explainable, safety-critical autonomous systems. Its scalable design supports diverse sensor setups and complex environments, promising broad industry impact and safer deployment of autonomous vehicles.
Technical Contribution
The paper introduces a novel BEV-centric neuro-symbolic framework that explicitly models multi-source perception coordination via the ICA module, and enforces entity-grounded verification through the SSRE engine. This combination enables deterministic conflict resolution, provenance tracking, and fact-based reasoning, breaking away from black-box deep models. The architecture supports multi-task scene understanding, improves robustness against sensor noise, and offers a transparent decision pipeline, setting new benchmarks for safety and interpretability in autonomous driving.
Novelty
This is the first work to embed explicit multi-source information coordination and symbolic verification within an autonomous driving perception-reasoning pipeline. Unlike prior approaches that treat perception and reasoning separately, it constructs a unified, auditable SceneSummary, integrating neural perception with formal logical validation. This innovative combination effectively reduces hallucinations and enhances trustworthiness, establishing a new paradigm for safe autonomous cognition.
Limitations
- The system's robustness under extreme weather or sensor failure conditions remains limited, as perception errors can still propagate into reasoning.
- Computational overhead from multi-source fusion and verification processes poses challenges for real-time deployment, requiring further optimization.
- Handling highly complex or ambiguous scenarios with multiple conflicting sources needs more adaptive conflict resolution strategies, which are future directions.
Future Work
Future efforts will focus on integrating end-to-end learning for perception and reasoning, reducing computational costs, and enhancing robustness against adverse conditions. Exploring adaptive conflict resolution and uncertainty modeling will further improve system reliability. Extending the framework to multi-vehicle cooperation and edge deployment could unlock broader applications, pushing autonomous systems toward safer, more explainable, and more trustworthy operation.
AI Executive Summary
Autonomous driving relies heavily on accurate scene understanding, yet existing perception systems often struggle with inconsistent sensor data and hallucinations, especially in complex environments. Traditional deep learning models excel at feature extraction but lack mechanisms for ensuring the semantic consistency and verifiability of high-level reasoning. This gap poses safety risks and limits trustworthiness.
To address this, the authors propose InfoCoordiBridge, a novel neuro-symbolic architecture centered on Bird’s-Eye-View (BEV) perception. The system integrates multiple sensors—cameras, LiDAR, radar—via a multi-agent perception layer that outputs structured facts and modality summaries. An ICA module then performs deterministic entity alignment, coordinate normalization, and conflict-aware attribute fusion, creating a unified SceneSummary that serves as a reliable, auditable representation of the environment.
Building on this, the SceneSummary-Driven Reasoning Engine (SSRE) employs a chain-of-prompt verification workflow, grounded in entity-based reasoning, to ensure factual consistency and reduce hallucinations. This approach effectively filters out conflicting or redundant perception evidence before high-level decision-making. Experimental results on nuScenes and Waymo datasets demonstrate that the system maintains competitive 3D detection accuracy while boosting fusion consistency to 98%, with redundancy below 1%. QA tasks further show a 20% reduction in hallucinated entity mentions and improved factual grounding.
Overall, this work significantly advances autonomous scene understanding by explicitly modeling perception conflicts and embedding verification into the reasoning process. It offers a scalable, interpretable framework that enhances safety, trustworthiness, and explainability, paving the way for more reliable autonomous vehicles. Future directions include optimizing computational efficiency, extending robustness, and integrating adaptive conflict resolution strategies for even safer deployment.
Deep Analysis
Background
近年来，自动驾驶技术在多模态感知融合方面取得了显著进展。代表性方法如BEVFusion、TransFusion等，强调在鸟瞰图空间实现多源特征的对齐，显著提升了3D目标检测性能。然而，这些方法多局限于几何和低层次语义，难以满足高层场景理解的需求。随着大规模语言模型（LLMs）的兴起，将自然语言理解引入自主驾驶成为解决复杂交互和推理的关键途径。尽管如此，现有系统多采用后置LLM推理，缺乏主动协调感知冲突和验证机制，导致幻觉和信息不一致问题频发，影响系统的安全性。近年来，符号推理与深度学习结合的神经符号架构逐渐受到关注，旨在提升推理的可解释性和可靠性，但在自主驾驶中的应用仍处于探索阶段。
Core Problem
核心问题在于如何在多模态感知基础上，建立一个具有可追溯性和验证能力的场景理解框架。现有方法多忽视感知信息的冲突与不一致，导致推理结果缺乏可信度。在复杂交通环境中，传感器噪声和环境干扰引发的感知误差，容易被模型误判为真实实体，造成幻觉。解决这一问题，需设计一种机制，能够在信息融合阶段进行冲突检测与修正，并在推理阶段引入验证流程，以确保决策的安全性和可信性。
Innovation
本研究的创新点包括：1）提出以BEV为核心的神经符号架构，建立多源信息的显式协调机制，确保信息一致性；2）设计ICA模块，实现实体对齐、冲突融合和可追溯的场景表示，增强系统的透明性；3）开发SSRE推理引擎，结合验证驱动的推理链，显著减少幻觉实体，提升推理的真实性和安全性。这一架构突破了传统深度模型的黑箱限制，将符号推理引入感知-决策流程，有效缓解了感知冲突和虚假信息的问题，为自主驾驶提供更安全、可信的认知基础。
Methodology
- �� 多智能体感知层：利用BEVFusion、LiDAR、相机和雷达模型，输出结构化事实（JSON格式）和模态摘要。
- �� ICA模块：对多源实体进行坐标归一化、匹配和属性融合，采用协方差交叉融合、置信度加权和一致性门控，生成SceneSummary。
- �� SceneSummary：包含实体状态、属性、源信息和不确定性指标，确保信息的可追溯性。
- �� SSRE推理引擎：基于验证驱动的链式推理，结合实体基础的推理流程（如SSP、SRA、VRCG、IVL），实现推理的真实性验证。
- �� 实验：在nuScenes和Waymo数据集上，比较不同模块的性能，评估融合一致性、虚假实体比例和问答准确率，进行消融分析。
Experiments
采用nuScenes和Waymo公开数据集，设置多模态感知模型、融合算法和推理验证流程。指标包括3D检测AP、融合一致性、虚假实体比例和问答准确率。对比BEVFusion、TransFusion、DriveAgent等，验证ICA和SSRE的提升效果。通过不同场景（如夜间、雨天）进行鲁棒性测试，分析系统在复杂环境下的表现。实验还包括消融研究，验证各模块贡献。
Results
系统在nuScenes和Waymo上保持了约70%的3D检测AP，融合一致性提升至98%，冗余率低于1%。QA任务中，虚假实体提及率降低20%，事实准确率提升15%。消融实验显示，ICA模块对冲突解决至关重要，SSRE显著减少幻觉实体。整体效果优于现有主流方法，验证了信息协调与验证机制的有效性。
Applications
该架构适用于自动驾驶车辆的场景理解、决策支持和安全监控。依赖多模态传感器输入，结合符号推理，提升系统的可信度和可解释性。可应用于城市交通、自动泊车和无人配送等场景，为行业提供安全、可靠的认知基础。未来还可扩展到多车协作和边缘计算环境，增强系统的适应性。
Limitations & Outlook
当前模型对极端天气和复杂交互场景的鲁棒性仍需提升，感知误差可能影响推理准确性。融合与验证过程计算成本较高，实时性待优化。多源冲突处理在极端情况下仍存在不足，未来需引入自适应机制和更高效的推理验证技术。
Plain Language Accessible to non-experts
想象你在厨房里准备一顿大餐。每个厨师负责不同的菜肴，比如一个负责炒菜，一个负责煮汤，还有一个负责摆盘。每个人都用不同的方法和材料，但都要确保菜肴协调一致，味道好。这个过程就像自动驾驶中的多传感器——每个传感器像厨师一样，提供不同的“菜肴”信息。为了做出一盘完美的菜肴，你需要把这些信息整理、核对，确保没有冲突或错误。本文提出的系统就像一个厨房管理者，能把所有厨师的菜肴信息协调好，确保每道菜都正确无误，然后再把这些信息用在“决策”这个“菜肴”上。这样，自动驾驶的“厨师们”就能合作得更好，做出安全又美味的“菜肴”。
ELI14 Explained like you're 14
想象你在学校里参加一个团队项目，每个人都带来了自己的想法和资料。有的人带了图片，有的人写了报告，还有的人说了很多建议。有时候，这些资料会出现不一致，比如有人说某个地方很危险，有人说很安全。这时候，你需要一个聪明的朋友帮你整理这些信息，把正确的部分筛出来，冲突的地方解决掉，然后告诉你最靠谱的方案。这个朋友就像论文里的系统，它能把来自不同传感器的“信息”整理成一个清晰的“故事”，确保没有误会或虚假的信息。这样，老师（自动驾驶系统）就能根据这个“故事”做出正确的决定，保证每次出行都安全顺利。这个系统就像一个超级聪明的整理者，帮你把复杂的资料变得简单又可靠。
Abstract
Reliable autonomous driving requires scene understanding that is semantically consistent across heterogeneous sensors and verifiable at the reasoning stage. However, many recent LLM-driven driving systems attach the language model as a post-processor and force it to reason over redundant or conflicting perception outputs, which can amplify hallucinated entities and unsafe conclusions. This paper proposes InfoCoordiBridge, a BEV-centric neuro-symbolic architecture that inserts an explicit coordination bridge between perception and language reasoning. InfoCoordiBridge comprises (i) a unified multi-agent perception layer that outputs typed structured facts together with modality-focused synopses, (ii) an ICA module that aligns and fuses multi-source outputs into a single SceneSummary, and (iii) an SSRE module that performs SceneSummary-grounded reasoning with verification. Experiments on nuScenes and Waymo show that ICA preserves competitive 3D detection accuracy while substantially improving fusion consistency, reducing redundancy to below 1% and achieving about 98% attribute agreement. On NuScenes-QA and a template-aligned Waymo-QA benchmark, SSRE improves factual grounding and reduces hallucinated entity mentions compared with representative VLM and agentic baselines. Overall, by coordinating multi-sensor outputs into a single conflict-aware SceneSummary before prompting, InfoCoordiBridge prevents redundant and cross-modally inconsistent perception evidence from propagating into high-level reasoning.