Large Language Models and Agentic AI Architectures for Aerial and Ground Robots: A Systematic Review, Taxonomy, and Design Guidelines
This systematic review analyzes 124 studies to establish a taxonomy of six architectural patterns for integrating large language models and agentic AI into aerial and ground robots, offering design guidelines that prioritize task dynamics, safety, and deployment constraints over benchmark scores while highlighting current limitations in onboard inference and physical validation.
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 robot that can understand a simple sentence like "find the lost dog in the park" and then figure out the steps to do it on its own. For decades, robots have been excellent at following strict, pre-written instructions, but they struggle when the world changes or when asked to do something they haven't been explicitly programmed to do. Recently, a new type of computer brain, known as a large language model, has emerged. These are the same powerful systems that can write stories, answer questions, and hold conversations. Scientists have begun trying to connect these language brains to physical robots, hoping to give machines the ability to reason, plan, and adapt just like a human would. This fusion of language understanding and physical movement is the frontier of modern robotics, promising machines that can work in unpredictable environments, from disaster zones to busy city streets.
A researcher at the University of East London, Abdelrahman Mahmoud, recently took a deep look at how this is actually happening. He did not build a new robot himself; instead, he acted as a cartographer for a rapidly expanding and somewhat chaotic landscape. He gathered and analyzed 124 different scientific studies published between January 2020 and April 2026, covering both flying robots, known as drones, and ground-based vehicles. His goal was to map out the different ways engineers are currently connecting these language brains to robot bodies and to understand which methods work best for which jobs. The result is a clear picture of a field that is moving away from the idea of a single, all-powerful robot brain and toward a more practical, team-based approach.
The study found that the most successful robots today do not rely on one giant model to do everything. Instead, they use a hybrid system where different parts of the robot handle different tasks. The language model acts as the high-level commander. It listens to a human's request, breaks it down into smaller steps, and decides what needs to be done next. However, when it comes to the actual physical work, the robot relies on older, more specialized tools. For example, while the language model might decide that the robot needs to "look for a red door," a separate, highly specialized computer vision program is what actually identifies the red door in the camera feed. Similarly, while the language model might plan a route, a traditional flight controller or wheel controller is what keeps the robot from crashing into a wall or falling off a cliff. The research shows that trying to use the language model to do the precise, split-second math required for movement is currently too slow and too prone to errors.
Mahmoud identified six main ways these systems are built, ranging from simple to complex. The simplest design is a one-way street: the robot sees something, the language model makes a plan, and the robot executes it without looking back. This works well for easy tasks in quiet rooms, but if the robot bumps into something unexpected, it gets stuck. More advanced designs add a feedback loop, allowing the robot to check its work and change its plan if things go wrong. The most sophisticated systems use a hierarchy, where a central language model acts as a manager, delegating specific jobs like "look here" or "move there" to specialized sub-teams. Some researchers are even testing systems where multiple robots talk to each other, each with its own language-based brain, to solve problems together. Despite the excitement around new, all-in-one models that try to see, think, and move all at once, the review found that these are still rare and often less reliable than the hybrid approach.
The analysis also highlighted a significant gap between what robots can do in a computer simulation and what they can do in the real world. Out of the 124 studies reviewed, only 44 actually tested their systems on physical robots. The rest relied on simulations, which are useful for testing ideas but often hide the messy, unpredictable problems of the real world. Furthermore, the study noted that while these robots are getting smarter, they still struggle with things humans find easy, like understanding exactly where an object is in three-dimensional space or planning a long sequence of actions without making a mistake. They can sometimes "hallucinate," inventing objects or steps that do not exist, which is why safety checks and human oversight remain essential.
Ultimately, this review suggests that the future of intelligent robots lies not in replacing human engineers with a single magical brain, but in building better teams. The most effective robots appear to be those that combine the flexible, creative reasoning of language models with the precise, reliable speed of traditional robotics software. The flying robots in the study focused heavily on looking at large areas from above, while the ground robots focused on interacting with objects and people nearby, yet both followed the same basic rule: let the language model decide the "what" and the "why," but let the specialized, proven systems handle the "how." As this technology matures, the challenge will be to make these systems faster, safer, and more capable of working in the complex, chaotic world we live in, without losing the reliability that keeps them from causing harm.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.