TLDR: A new research paper introduces the Self-organizing Nervous System (SoNS) which enables robot swarms to automatically request and execute new code generated by a Large Language Model (LLM) when they encounter obstacles. This system allows the swarm to be programmed as a single entity and provides global environmental awareness to the LLM, leading to an 85% mission success rate in adapting to unforeseen challenges in both real-robot and simulated environments.
Imagine a group of robots working together, but when they hit an unexpected snag, they can instantly ask an artificial intelligence for new instructions and learn on the fly. This is precisely what new research from Weixu Zhu, Marco Dorigo, and Mary Katherine Heinrich at IRIDIA, Universit´e libre de Bruxelles, demonstrates with their innovative approach to robot swarms.
Traditionally, programming robot swarms for complex group behaviors has been a significant challenge. Robots are programmed individually, but the desired outcome is a collective behavior. This often leads to extensive trial-and-error testing because designing these self-organized group behaviors can be incredibly difficult to predict analytically. Furthermore, robots in a swarm typically only have local information about themselves and their immediate neighbors, making it hard for the entire swarm to assess its global performance or configuration.
The researchers propose that their recently introduced Self-organizing Nervous System (SoNS) can dramatically simplify the implementation of online automatic code generation for robot swarms using Large Language Models (LLMs). SoNS acts as a kind of middleware, allowing robots to form and dissolve temporary, centralized control structures in a self-organized manner. This means a swarm can coordinate its collective sensing, actuation, and decision-making as if it were a single, reconfigurable robot, without losing the benefits of scalability, flexibility, and fault tolerance inherent in self-organization.
The SoNS offers two key advantages for integrating LLMs:
Ease of Behavior Design
Because the SoNS creates a hierarchical structure that separates global and local actions, an LLM can directly provide code for a desired global behavior. This is a significant improvement over trying to design individual robot behaviors that would, through complex interactions, eventually lead to the desired group outcome.
Also Read:
- AutoMaAS: A Self-Evolving Framework for Multi-Agent AI Systems
- AI Models Learn to Adapt and Specialize with Self-Curated Training
Global Estimation of Swarm and Environment
With SoNS, sensor information is forwarded upstream through the self-organized network to an interchangeable ‘brain robot’. This brain robot can then provide the LLM with a comprehensive estimate of the entire swarm’s configuration and its sensed environment, offering a global perspective that was previously difficult to achieve.
In proof-of-concept demonstrations, the researchers showed that a SoNS-enhanced robot swarm could automatically solicit and run code generated online by an external LLM. The mission was simple: move forward while maintaining a square formation. If the swarm got stuck due to obstacles, the current SoNS-brain robot would initiate a conversation with the LLM (specifically, DeepSeek R1 via the OpenRouter API). It would send basic information about the swarm’s context, robot capabilities, mission goal, and current sensor data, along with a generic request for help. The LLM would then generate new code in Lua, which the brain robot would receive, update its program, and then disseminate to all other robots in the swarm in a self-organized manner.
The demonstrations included a real-robot setup with 6 robots (4 ground, 2 aerial) navigating 2 unknown obstacles, and 20 simulation trials with 34 robots (25 ground, 9 aerial) facing 15 unknown obstacles. The system achieved an impressive 85% mission success rate in simulations. While most trials were successful, some unsuccessful instances occurred when the LLM generated code that inadvertently removed mechanisms for detecting when other robots were stuck, leaving some behind. This highlights areas for future work, such as more stringent separation of static and updateable code, and more principled approaches to requesting help for complex tasks.
This research marks a significant step towards more adaptive and resilient robot swarms, capable of learning and evolving their behaviors in real-time to overcome unforeseen challenges. You can read the full research paper here: Online automatic code generation for robot swarms: LLMs and self-organizing hierarchy.
