Jadwal Sholat

Memuat jadwal sholat…

Fisika & Matematika editorial

Open AccessOA2023

Multi-Agent Deep Reinforcement Learning for Multi-Robot Applications: A Survey

This survey reviews multi-agent deep reinforcement learning (MADRL) for multi-robot systems, categorizing applications and discussing open challenges and future opportunities. It fills a gap since the last survey in this domain dates to 2004, before deep reinforcement learning existed.
James Orr; Ayan Dutta· Italian National Conference on Sensors· 2023· DOI 10.3390/s23073625

The core problem

Deep reinforcement learning (DRL) has achieved notable successes in fields such as mathematics, games, healthcare, and robotics. This paper focuses on multi-agent deep reinforcement learning (MADRL), where multiple agents in a shared environment learn not only from their own experiences but also from each other. The authors emphasize applications in multi-robot systems, where a single robot may be insufficient to complete a task, necessitating collaboration among multiple robots toward a common global objective. Despite the importance of MADRL for multi-robot systems from both theoretical and applied perspectives, the most recent survey in this domain dates to 2004—a time when DRL had not yet been invented. This survey aims to fill that gap by reviewing recent literature, classifying papers primarily by their multi-robot applications, discussing challenges, and suggesting future applications that could benefit from MADRL advances.

Innovation

The authors conduct a systematic literature review of MADRL applied to multi-robot systems. They classify the reviewed papers primarily based on their multi-robot applications. The survey covers works that employ MADRL algorithms—such as value-based, policy-gradient, and actor-critic methods—in multi-robot settings. The classification framework organizes the literature according to application domains (e.g., exploration, transportation, surveillance) and learning paradigms (e.g., centralized training with decentralized execution). The authors also discuss challenges faced by current research and provide a potential list of future applications. The methodology emphasizes the importance of considering multi-agent learning dynamics, where agents learn from both individual and shared experiences, and the need for scalable and robust solutions in real-world multi-robot deployments.
Introduction
Deep reinforcement learning (DRL) has achieved notable successes in fields such as mathematics, games, healthcare, and robotics. This paper focuses on multi-agent deep reinforcement learning (MADRL), where multiple agents in a shared environment learn not only from their own experiences but also from each other. The authors emphasize applications in multi-robot systems, where a single robot may be insufficient to complete a task, necessitating collaboration among multiple robots toward a common global objective. Despite the importance of MADRL for multi-robot systems from both theoretical and applied perspectives, the most recent survey in this domain dates to 2004—a time when DRL had not yet been invented. This survey aims to fill that gap by reviewing recent literature, classifying papers primarily by their multi-robot applications, discussing challenges, and suggesting future applications that could benefit from MADRL advances.

Why it matters

The authors discuss several challenges in applying MADRL to multi-robot systems:

- **Scalability**: As the number of robots increases, the joint action space grows exponentially, making learning difficult.
- **Non-stationarity**: Each agent's policy changes during learning, creating a non-stationary environment for other agents.
- **Partial observability**: Robots often have limited sensing capabilities, requiring them to learn from incomplete information.
- **Sim-to-real transfer**: Policies learned in simulation may not transfer well to real robots due to modeling errors and noise.
- **Communication constraints**: Limited bandwidth and reliability affect coordination.

The authors also provide a potential list of future applications, such as environmental monitoring, infrastructure inspection, and autonomous warehouses. They emphasize the need for more real-world experiments and benchmarks to validate MADRL algorithms. The survey concludes that MADRL holds tremendous promise for multi-robot systems but requires further research to address the identified challenges.

Who should read this

Pembaca Research Digest dan praktisi rumpun ini.

Opening member content…