What behavior trees bring to robotic navigation
To understand what Perando's research contributes, we first need to briefly explain two key concepts: large language models in robotics and behavior trees.
Large language models (LLMs) — the technology behind ChatGPT, Gemini, or Claude — have been increasingly used in the last two years to control robots. The principle is relatively straightforward: the robot receives a verbal command ("go to the kitchen and avoid obstacles") and the LLM translates it into a series of specific actions that the robot executes step by step. The advantage is obvious — the robot doesn't need to be programmed for each specific situation; instead, it uses the general "understanding of the world" that the language model acquired during training. The problem, however, is that LLMs sometimes hallucinate, meaning they generate outputs that look meaningful but actually don't make sense. And what is at most annoying in a regular chat can be outright dangerous with a physical robot in real space.
This is where behavior trees come into play. They are a mathematical model describing how a system should decide between different tasks. Imagine it as a decision tree with a fixed structure: individual "leaves" of the tree represent specific actions (turn left, stop, continue straight) and branches determine in what order and under what conditions they should be performed. Behavior trees have been used for years in the video game industry to control non-player character behavior — titles like Halo, Bioshock, or Spore used them. Their key advantage? They are visually intuitive, easy to test and debug, and most importantly — their behavior is predictable and verifiable.
Perando and his advisor, professor Xunyu Pan (head of the FLAIR lab — Frostburg Laboratory for Artificial Intelligence and Robotics), proposed an approach called Constrained Behavior Tree Generation — that is, generating behavior trees with built-in constraints. In practice, this means that when an LLM proposes a series of steps for the robot, the system automatically validates each of these steps against a behavior tree that defines what is safe in the given situation. If an action does not lead to the intended goal or violates safety constraints, the system rejects it and looks for an alternative. Simply put: the robot has its own ideas approved before executing them.
From elementary school to IEEE — the story of Jeremy Perando
What is also remarkable about Jeremy Perando's story is how quickly he became a publishing researcher. As recently as 2025, he presented his first work — Enhancing Robotic Navigation with Large Language Models — which mapped out the basic possibilities of using LLMs in robotic navigation. A year later, he already stands as the first author of a paper at an international IEEE conference, a milestone typically reached by doctoral students.
The IEEE EIT 2026 (Electro/Information Technology) conference took place on May 21–23, 2026 in La Crosse, Wisconsin. It is an established conference that annually attracts researchers from the fields of electrical engineering, computer science, and robotics. The fact that an undergraduate student is the first author is, according to a statement by the FSU Computer Science Department, a historic milestone.
The development took place in the FLAIR laboratory under the guidance of Dr. Xunyu Pan and with support from the NSF Appalachian Innovation Institute (AI²) grant — an initiative of the U.S. National Science Foundation focused on supporting technological innovations in the Appalachian region, an area that traditionally does not rank among technology hubs. This adds another dimension to Perando's achievement — it shows that top-tier AI research doesn't have to emerge only from MIT or Stanford.
Why AI robot safety is a critical topic
The combination of large language models and physical robots is one of the most watched trends in AI research for 2026. Nvidia last year introduced the GR00T project, a universal foundation model for humanoid robots. Google DeepMind is intensively working on robots that combine language understanding with physical manipulation. The entire concept of "embodied AI" — that is, artificial intelligence that has a physical body and interacts with the real world — is experiencing an unprecedented boom.
But the more autonomy we give robots, the more urgent the question of their safety becomes. While for a chatbot, a hallucination means at worst a bad answer, for a robot it can mean running into a person, falling down stairs, or damaging equipment. LLM models were not designed with physical laws in mind, and their "understanding" of space is purely statistical — there is no guarantee that the proposed path is physically achievable.
The European Union is responding to this problem through the AI Act, whose provisions on high-risk AI systems directly affect autonomous robots as well. Manufacturers and operators of AI robots will have to demonstrate that their systems meet strict safety standards — including mechanisms for real-time decision verification. Approaches like constrained behavior trees may be one of the tools to demonstrate such compliance.
What this means for European and Czech robotics
At first glance, it might seem that research from a Maryland university is distant from the Czech environment. The opposite is true. The Czech Republic has a strong tradition in robotics — CTU in Prague operates the Czech Institute of Informatics, Robotics and Cybernetics (CIIRC), which ranks among the leading European robotics workplaces. BUT in Brno has long been dedicated to mobile robotics and industrial automation. Both institutions have increasingly focused in recent years on integrating artificial intelligence into robotic systems — and a safety layer such as constrained behavior trees could be a relevant research direction for them.
Moreover, the European approach to AI regulation creates specific demand for verifiable safety mechanisms. It's not enough to say "our robot is safe" — you have to be able to technically prove it. The tree structure of behavior trees is ideal for this purpose because it enables formal verification: each branch of the tree can be mathematically verified independently of others and guarantee that, under given conditions, the robot will execute only permitted actions.
Future: AI models directly in robots
The FLAIR research team has already announced the direction they will take next: they want to move AI model execution from the cloud directly into robots — onto the so-called edge. This is a fundamental shift. Currently, most LLM-controlled robots send data for processing to remote servers, which introduces latency (delay), dependency on internet connectivity, and last but not least, security risks associated with data transmission.
Running the model directly on the device — typically on specialized chips like Nvidia Jetson or Google Coral — would allow robots to make decisions in real time without cloud dependency. This is especially crucial for deployment in industrial facilities, warehouses, or healthcare settings, where every millisecond of latency can matter.
The combination of local inference and formally verified behavior trees represents, according to Pan and Perando, an architecture that could serve as the foundation for the next generation of safe autonomous robots. The NSF AI²-funded research will continue in this direction, and it is likely that we will hear more from the FLAIR lab.
What is the difference between a regular LLM robot and a robot with a behavior tree?
In the common approach, the LLM proposes a series of steps and the robot executes them — without additional checking. With a behavior tree, each proposed step goes through a validation structure that verifies whether the action is safe and whether it actually leads to the goal. If not, the system blocks the action and looks for another path. It's similar to the difference between believing everything someone says and having facts checked by an independent source.
Is the research usable for ordinary consumer robots, or only for laboratory conditions?
The principle of constrained behavior trees is universal and does not depend on a specific type of robot. It can be applied to robotic vacuum cleaners, delivery drones, industrial AGVs, and autonomous vehicles. The key is that safety constraints are defined for the specific environment and robot — a laboratory prototype and a commercial product use the same principle, differing only in the specific set of rules. With the approaching effectiveness of the EU AI Act, moreover, pressure is growing on manufacturers to integrate such mechanisms.
How do Czech universities approach this topic?
Czech technical universities — especially CTU (CIIRC) and BUT in Brno — have long-standing strong robotics programs. Integration of language models into robotics is still in an early phase here, but the combination of behavior trees and LLMs is a direction that naturally builds on existing expertise in industrial robotics, trajectory planning, and multi-agent systems. Similar research projects can be expected to appear in the Czech environment within one to two years.
Source: WV News — Frostburg student presents AI-driven robotics research at IEEE international conference
