From Simulation to Real Robots: Multi-Agent Reinforcement Learning for Task Assignment and Navigation

Introduction

Here we provide the additional materials to support the paper.

Real-World Deployment

In this section, we provide the details of the real-world deployment of FMAPPO on a fleet of mobile robots.

Real-World Deployment With Different Command Update Frequencies

Here the same model that was trained in simulation with command update frequency of 20 Hz is deployed in the real world with different command update frequencies. ALL VIDEOS ARE SHOWN IN REAL TIME SPEED, NO SPPED UP (1x).

First scenario with different command update frequencies (IN REAL TIME SPEED)
Command Update Frequency 20 Hz Command Update Frequency 40 Hz
Command Update Frequency 60 Hz Command Update Frequency 120 Hz
Second scenario with different command update frequencies (IN REAL TIME SPEED)
Command Update Frequency 20 Hz Command Update Frequency 40 Hz
Command Update Frequency 60 Hz

Real-World Deployment With Action Smoothing

Here the same model that was trained in simulation with command update frequency of 20 Hz is deployed in the real world with command update frequency of 40 Hz and action smoothing by averaging the actions over a window of size 5. THE VIDEOS IS SHOWN IN REAL TIME SPEED, NO SPPED UP (1x)

Contact

Comming Soon