Empowering Internet of Vehicles with Embodied Intelligence: A Survey of Technologies, Evaluation, and Outlook
This survey presents a comprehensive framework for empowering the Internet of Vehicles with embodied intelligence by systematically analyzing the closed-loop integration of perception, decision-making, and control, while evaluating existing technologies, datasets, metrics, and future challenges to enhance vehicle adaptability and safety.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a car that does not just see the road ahead, but understands the world around it in the same way a human driver does. It does not merely process a stream of data points; it feels the flow of traffic, anticipates the hesitation of a pedestrian, and adjusts its speed based on the subtle shift of a vehicle merging from a blind spot. This is the promise of the Internet of Vehicles, a system where cars, traffic lights, and roadside sensors talk to one another to create a shared picture of the road. Yet, for all this connectivity, current systems often operate in a broken loop. They perceive a hazard, make a decision, and act, but they rarely use the result of that action to immediately correct their understanding. If a car misjudges a gap and brakes too hard, the system might not realize its error until the next cycle, allowing mistakes to pile up like uncorrected notes in a song. The gap between seeing the world and truly understanding it in real-time remains a significant hurdle for safe, autonomous travel.
A new survey by researchers from Guizhou University and their international colleagues seeks to bridge this gap by introducing the concept of embodied intelligence to the Internet of Vehicles. Rather than treating perception, thinking, and driving as separate steps in a factory line, the authors argue that these functions must form a continuous, closed loop. In this view, a vehicle is an "embodied agent," a physical entity that constantly interacts with its environment. It sees, it thinks, it acts, and then it immediately uses the feedback from that action to update its understanding of the world. This cycle repeats thousands of times a second, allowing the vehicle to correct errors on the fly and adapt to the chaotic, shifting reality of a busy street. The survey, posted on November 9th, 2026, maps out the technologies required to build this loop, from how cars share information to how they learn from their mistakes, offering a roadmap for moving from simple data sharing to genuine, adaptive cooperation on the road.
The researchers began by organizing the vast landscape of existing technology into four distinct but connected stages: perception, cognition, decision-making, and execution. In the first stage, perception, the survey details how vehicles are moving beyond their own sensors to build a complete picture of the road. Instead of relying solely on a camera or radar mounted on the car, modern systems are learning to combine data from multiple sources, including cameras on other cars and sensors on streetlights. This cooperative perception allows a vehicle to "see" around corners and through heavy traffic, filling in the blind spots that have long plagued autonomous driving. The survey highlights how these systems align data from different angles and times, stitching together a unified view of the road that is far more reliable than any single car could achieve alone.
Once the vehicle has a clear view of the world, it must understand what is happening, a stage the authors call cognition. Here, the vehicle moves from simply detecting objects to interpreting their intentions and the rules of the road. It asks not just "what is that?" but "what will that do next?" and "what does the traffic law say?" The survey reviews how vehicles are learning to predict the behavior of pedestrians, cyclists, and other drivers, using historical data and shared information to guess future movements. This stage also involves memory; the vehicle recalls past interactions and learns from them, building a mental model of the traffic environment that evolves as the journey continues. This cognitive layer is what allows a car to understand that a driver signaling a turn is likely to move, or that a group of pedestrians waiting at a curb is about to cross.
With a solid understanding of the scene, the vehicle must then decide what to do. The survey explains that decision-making in the Internet of Vehicles is not just about choosing a path for one car, but about negotiating with others. In a connected environment, vehicles can coordinate their moves, agreeing on who goes first at an intersection or how to merge smoothly into a tight gap. The authors examine how these decisions are made, ranging from rule-based systems that follow strict traffic laws to advanced learning systems that adapt to complex, unpredictable situations. The key finding here is that effective decision-making requires a constant dialogue between the vehicle and its surroundings, ensuring that the chosen action is not only safe but also feasible given the physical limits of the car and the behavior of others.
Finally, the vehicle must execute its decision, turning a plan into physical motion. This execution stage is where the rubber meets the road, literally. The survey details how control systems translate high-level decisions into precise steering, acceleration, and braking commands. It emphasizes that this step must be tightly coupled with the earlier stages; if the car drifts slightly off course, the system must immediately sense this deviation and adjust its understanding and plan accordingly. The researchers highlight the importance of safety mechanisms that act as a final check, ensuring that even if a decision is flawed, the physical execution remains within safe limits. This closed-loop feedback, where the result of an action instantly informs the next thought, is the core of embodied intelligence.
The survey also takes a hard look at the tools available to researchers, reviewing the datasets and metrics used to test these systems. It points out that while there are many simulations and test cases, there is a shortage of real-world data that captures the full complexity of cooperative driving. The authors note that current evaluations often focus on isolated tasks, such as detecting a single object, rather than the entire cycle of seeing, thinking, and acting. They argue that to truly advance, the field must move toward testing how well these systems perform in continuous, real-world loops, where errors in one stage can cascade into failures in the next. The paper suggests that future progress will depend on developing better ways to handle uncertainty, predicting the intentions of others across different cities and cultures, and learning from rare, difficult driving scenarios that do not appear often in standard training data.
Ultimately, this survey serves as a comprehensive guide to a shifting paradigm in autonomous driving. It moves the conversation away from simply connecting cars to the internet and toward creating vehicles that are truly alive to their environment. By framing the vehicle as an embodied agent in a continuous loop of perception and action, the authors provide a clear path forward for researchers and engineers. The work suggests that the future of intelligent transportation lies not in faster computers or more sensors alone, but in the seamless integration of these elements into a system that learns, adapts, and corrects itself in real-time. As the technology matures, the goal is to create a driving experience that is not just automated, but genuinely intelligent, capable of navigating the unpredictable human world with the same fluidity and awareness as a skilled driver.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.