Technical Summary: Multi-Agent Embodied Autonomous Driving: From V2X Information Exchange to Shared World Models

1. Problem Definition

The paper addresses the transition of autonomous driving (AD) from isolated, single-agent intelligence to Multi-Agent Embodied Autonomous Driving (MAEAD). While traditional AD focuses on robust, self-contained vehicles, this paradigm is constrained by partial observability, sensor occlusion, and the inability to reason about dense social interactions.

Current approaches largely rely on Vehicle-to-Everything (V2X) Information Exchange, where agents broadcast raw data, processed features, or high-level intentions (e.g., BSMs, CAMs). The authors argue that this "Information Exchange" paradigm is insufficient for dense, uncertain scenarios because it treats agents as loosely coupled decision-makers that merely correct local world views. It fails to achieve the mutual predictability, coordinated intent, and joint future reasoning required for safe, complex traffic. The core problem is how to evolve from simple data sharing to a system where agents co-construct and maintain a Shared World Model (SWM) that aligns state, intent, and action across the fleet.

2. Methodology and Framework

The paper is a structured narrative review synthesizing over 380 publications. It does not propose a new algorithm but establishes a system-level evidence chain to evaluate existing literature.

A. The Organizing Lens: Shared World Models (SWMs)

The survey uses the transition from Information Exchange to SWM-driven coordination as its primary lens. An SWM is defined as a dynamic, predictive representation of the environment and its agents, maintained across the fleet to support collective decision-making.

B. Operational Evidence Criteria

To categorize and compare heterogeneous methods, the authors define three operational criteria (C1–C3) that a method must satisfy to be considered part of the SWM-driven progression:

• C1: Cross-agent State Alignment: The method must maintain an explicit shared or aligned state representation (e.g., fused BEV features, shared latent maps) rather than just broadcasting raw detections.
• C2: Intent and Plan Alignment: The method must explicitly model, communicate, or negotiate future behavior, goals, or roles, moving beyond reactive control.
• C3: Coordinated Downstream Action: The shared state or modeled intent must directly influence joint policy selection, trajectory generation, or multi-agent control, closing the loop between perception and action.

C. Scope of Review

The review covers the following technical domains, analyzing how they contribute to the evidence chain:

1. Collaborative Perception (CP): Examining Early, Intermediate, and Late fusion paradigms.
2. Inter-Agent Cognition & Communication: Analyzing explicit V2X protocols, implicit behavioral signaling, semantic communication, and LLM-based negotiation.
3. Collaborative Decision-Making & Planning: Reviewing game-theoretic approaches, Multi-Agent Reinforcement Learning (MARL), and generative/world-model-based planning.
4. End-to-End (E2E) Driving: Assessing systems that map shared latent states directly to control.
5. Simulation & Data Engines: Evaluating the fidelity of simulators and datasets used for validation.

3. Key Contributions

The paper's primary contributions are the synthesis of the field under a unified architectural narrative and the identification of specific research gaps:

1. Holistic Integration: Unlike prior surveys that focus on single layers (e.g., only perception or only communication), this work integrates collaborative perception, decision-making, foundation models, and validation infrastructure into a single framework.
2. SWM-Driven Coordination Lens: The authors operationalize the shift from data exchange to shared world modeling through the C1–C3 criteria, providing a standardized way to evaluate whether a method achieves true coordination.
3. Evidence Gap Analysis: The survey identifies that while shared-state construction (via intermediate fusion) is mature, intent alignment and coordinated downstream action remain underdeveloped, particularly regarding real-world safety guarantees.
4. Future Research Agenda: The paper outlines four critical gaps for future research:
   • Scalability: How MAEAD systems scale with agent density and communication budgets.
   • Open-Set Coordination: Handling legacy vehicles and non-cooperative human drivers without shared protocols.
   • Trustworthiness & Security: Defending against consensus-level attacks and malicious agents that poison shared beliefs.
   • Embodied Social Compliance: Ensuring systems adhere to implicit social norms and cultural driving behaviors.

4. Results and Findings

The survey synthesizes findings across the technical spectrum:

• Perception: Intermediate fusion (feature-level) is the dominant research direction, balancing accuracy and bandwidth better than Early (raw data) or Late (object-level) fusion. However, most methods remain focused on perception accuracy rather than downstream planning impact.
• Communication: Explicit V2X messages (BSMs, CPMs) provide machine-readable state but lack semantic richness. Semantic communication and LLM-based negotiation are emerging to make intent more inspectable, but they currently lack verified real-time safety guarantees and suffer from latency.
• Planning: Game-theoretic and MARL methods formalize interaction but struggle with credit assignment and transferability. Generative models (Diffusion, World Models) offer promising ways to simulate counterfactual futures, yet their integration into closed-loop, safety-critical control remains largely experimental.
• Validation: Evaluation is heavily concentrated in simulation (e.g., CARLA, V2X-Sim) and curated benchmarks. There is a significant lack of verified real-time safety guarantees in open traffic. The "simulation gap" regarding human behavioral realism is a major bottleneck.
• Foundation Models: LLMs and VLMs show promise for semantic grounding and reasoning but are currently best suited as advisory components alongside conventional safety filters, rather than sole decision-makers.

5. Significance and Claims

The paper claims that the field is at a critical juncture where the focus must shift from improving individual perception scores to verifiable shared-state maintenance and safe coordinated action.

• Modest Claims on Deployment: The authors explicitly state that current evidence for foundation-model-based coordination is simulation-centered. They do not claim that LLMs or world models are ready for deployment in open traffic.
• The "Audit" Requirement: The paper argues that for MAEAD to be deployable, systems must provide an auditable safety case. This includes logging raw messages, semantic claims, predicted trajectories, and rejected alternatives to allow for peer auditing of coordinated maneuvers.
• Hybrid Architectures: The authors suggest that the most viable path forward is hybrid architectures that combine rule-based safety constraints, MARL adaptation, and foundation-model-based reasoning. This approach aims to preserve interpretability and safety checks while leveraging the flexibility of learned models.

In conclusion, the paper posits that the next stage of autonomous driving research depends on moving beyond isolated agents to achieve verifiable shared-world-model alignment and safe coordinated action under real-time, lossy, and open-set traffic conditions.