Abstract
In this paper, we propose a new collision avoidance method among multiple autonomous mobile robots. Generally, it is difficult to apply conventional collision avoidance methods to robots in multi-robot environment because those methods are only applicable to a single robot environment. For this problem, we propose a new method using the LOCISS (LOcally Communicable Infrared Sensory System) by which a robot is able to communicate with other robots locally to exchange information necessary for their mutual collision avoidance. To realize the collision avoidance based on these information, we introduce a learning method by which a robot is able to acquire the behavior adaptively and autonomously. A learning curriculum is divided into multiple layers to reduce a number of situation for the learning. Finally, we implemented this system to execute acquired behaviors, so that robots can avoid other robots and obstacles at the same time. It is confirmed by conducting experiments that this method is effective in the multi-robot environment.
Introduction
In the field of swarm robotics, a major hurdle is collision avoidance in crowded spaces. Traditional methods designed for single robots often fail in multi-robot settings for two main reasons: sensor interference (crosstalk) from active sensors like ultrasonic rangefinders, and the high computational cost of passive sensors like vision when distinguishing between static and dynamic objects. Furthermore, Reinforcement Learning (RL) approaches often struggle with the "curse of dimensionality" in multi-robot environments, where the state space grows exponentially with the number of agents.
LOCISS: Locally Communicable Infrared Sensory System
The authors address these issues with the LOCISS hardware innovation. Instead of merely detecting obstacles, robots use LOCISS to exchange state data locally. Each robot is equipped with eight IR units providing 360-degree coverage.
This system serves a dual purpose:
- Obstacle Detection: IR signals reflecting off static objects like walls identify them as obstacles.
- Robot Communication: When a signal is received by another robot, they exchange IDs, speeds, and directions. This local "handshake" allows robots to negotiate trajectories without a central controller.
Multi-Layered Reinforcement Learning
To make learning tractable, the collision avoidance task is broken into a four-layer curriculum:
- Layer 1: Basic movement toward a goal.
- Layer 2: Avoiding a single object (wall or robot).
- Layer 3: Avoiding multiple objects within a specific sensor group.
- Layer 4: Comprehensive avoidance in complex environments.
By using this layered approach, the authors reduced the total number of states from a theoretical $5.7 \times 10^{13}$ to approximately $1.4 \times 10^5$, a reduction of eight orders of magnitude. This made it possible to store the entire behavior library in just 4.5MB of memory.
Experimental Results and Insights
The system was validated in both custom simulations and physical experiments using four omni-directional robots in a $4.0\text{m} \times 4.0\text{m}$ arena.
- Learning Efficiency: Stable avoidance behaviors converged within approximately 800 trials in simulation.
- Hardware Validation: Physical robots successfully negotiated paths in real-time during "cross-over" scenarios, adjusting trajectories and speeds to avoid both each other and the arena walls.
- Decentralized Decision Making: The system operates without a global map or GPS, making it highly scalable as each robot makes decisions based solely on its local neighborhood.
Conclusion
The integration of local communication via LOCISS and multi-layered RL provides a robust, decentralized solution for multi-robot navigation. While future work may address communication density in extremely crowded environments and autonomous curriculum design, this method effectively enables robots to navigate complex, dynamic environments by understanding the identity and intent of their neighbors.
Summary updated