Intelligent Optimization Model for Autonomous Systems Integrating Reinforcement Learning

View at publisher Summarise with VIRAT
Authors: Xuyao Zhao
Cited by: 0
References: 10

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This document presents an approach for the collective scheduling and the real-time avoidance of obstacles for the use of multiple mobile robots in a dynamic, heavy-duty setting within an intelligent warehousing and logistics scenario to aid in the fulfilment of the complex objectives of improving the throughput and robustness of the overall system. The study develops a collaborative decision making model that is based upon Multi-Agent Deep Deterministic Policy Gradient (MADDPG). A multi-objective reward function that combines the efficiency and safety constraints associated with task performance has been developed using a “centralized training and distributed execution” framework. The model also contains an agent interaction model that incorporates both a local observation space and a discrete action space. The proposed model has been compared and validated against six state-of-the-art algorithms (MAPPO (Multi-Agent Proximal Policy Optimization), QMIX, VDN (Value Decomposition Networks), IMPALA, SAC (Soft Actor-Critic), and IDQN) using simulators for eight different robot scale configurations (5-40 robots). It has been experimentally demonstrated that the proposed algorithm attains optimal results in a 10 -robot setting with average order completion time equal to 105.7 seconds, system throughput rate equal to 34.1 orders/hour, collision rate equal to 0.8 occurrences per 1000 movements, and success rate equal to 98.5 %. In addition, the results indicate that for high-density robotic configurations (40 robots), the amount of time required to complete orders was reduced by 17.1 % when compared to using the suboptimal algorithm. These findings support the ability of the MADDPG framework to provide an adequate solution for both path conflict and efficiency enhancement issues for multi-robot operations, through centralized assessment of value, which allows for the development of cooperative approaches among multiple robots.

Fields of Study & Subjects

Reinforcement Learning in Robotics Advanced Manufacturing and Logistics Optimization Robotic Path Planning Algorithms

Publication details

Journal / source
International Conference on Electrical Automation and Artificial Intelligence (ICEAAI)
Pages
918–922
Publication date
16 January 2026
Language
EN