Accessible Hardware Implementation for Multi-Agent Collective Construction

Martin Rames, Pavel SurynekAAAI 2025 · DOI

Abstract

Multi-Agent Collective Construction (MACC) research often suffers from a "reality gap" between idealized software simulations and expensive, specialized hardware. This paper presents MACC-SLIM (Simple Line-following Intelligent Machines), a low-cost platform using commodity Ozobots and an LCD screen to bridge this gap. By simulating 3D structures visually while maintaining physical robot dynamics, MACC-SLIM reveals critical synchronization delays and bottlenecks ignored in theoretical models.

Introduction

Multi-Agent Collective Construction (MACC) involves groups of robots collaborating to build complex structures. While theoretical models have advanced significantly, physical implementation remains a major hurdle. The current gold standard, the TERMES system, relies on custom-built robots and specialized blocks that are expensive and difficult to manufacture and maintain.

The TERMES system uses specialized hardware and blocks, creating a high barrier to entry for MACC research.

Due to these barriers, most researchers rely on software simulations. However, these simulations often assume "perfect" conditions—instantaneous movement, zero-time turns, and flawless synchronization—leading to a "reality gap" where theoretical plans fail in physical deployment. The authors aim to provide a middle ground: a system that is cheaper than custom hardware but more realistic than a pure simulation.

The MACC-SLIM Platform

The authors introduce the Simple Line-following Intelligent Machines (SLIM) system as a middle ground between pure simulation and expensive custom hardware.

Hardware and Interface

Instead of custom robots, SLIM utilizes Ozobots, small commodity differential-drive robots. The construction site is replaced by a flat-laid LCD screen which serves as both the floor and the controller. The screen displays animated colored circles and lines that the Ozobots follow and decode into high-level commands, such as "move forward one block" or "turn 90°."

2D-to-3D Mapping

To eliminate the need for physical blocks, the system simulates the 3D aspect visually. The screen displays the current height of the structure at each grid cell. While robots move in 2D, the system tracks their "virtual" height and handles the logic of block placement and retrieval. This allows researchers to focus on complex coordination problems without the mechanical overhead of physical manipulation.

The MACC-SLIM system coordinating multiple Ozobots on an LCD screen grid to build a benchmark structure.

Key Insights and Results

The researchers evaluated SLIM using benchmark structures and optimal plans generated via Mixed-Integer Linear Programming (MILP).

  • The "Turn Tax": Physical implementation revealed that turning takes significant time, a factor often ignored in models where "move" and "turn" are treated as equal timesteps. This leads to synchronization delays that can disrupt optimal plans.
  • Border Bottlenecks: SLIM highlighted that entry points to the construction zone are physical bottlenecks. Queuing at the border is a critical factor often overlooked in abstract simulations.
  • Scalability: The team successfully demonstrated 24 robotic agents collaborating to build a benchmark castle structure within a small laboratory space, proving the system's accessibility.

Conclusion

MACC-SLIM provides an accessible, low-cost "reality check" for MACC algorithms. By revealing the physical constraints of real-world robots, it allows researchers to refine theoretical models to better reflect physical reality. Future work will focus on integrating "turn-action" costs into pathfinding models and analyzing complex entry/exit interactions between the robot reservoir and the construction site.

Summary updated

BibTeX
@inproceedings{rames2025accessible,
  author = {Martin Rames and Pavel Surynek},
  title = {{Accessible Hardware Implementation for Multi-Agent Collective Construction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35368},
}
Figure