Assessment of Multi-Agent Reinforcement Learning Strategies for Multi-Agent Negotiation

In the realm of multi-agent systems, the effective coordination of agents for manipulation tasks poses a significant challenge. This study explores various strategies aimed at enhancing Multi-agent Reinforcement Learning (MARL) algorithms in the context of locomotion-based manipulation tasks. We systematically assess the performance of these strategies, incorporating reward shaping and algorithmic variations such as Proximal Policy Optimization (PPO) and Advantage Actor Critic (A2C). For better cooperation between agents, a prediction map is also implemented, informing the agents with a probability heat map of other agents based on their kinematic models. The experiments are conducted in the Isaac Sim simulation environment with Jetbot robots as agents. Results indicate distinct impacts of each strategy on critical performance metrics, including success rates, collision probabilities, and overall task efficiency. While some strategies exhibit notable improvements, others reveal limitations, emphasizing the nuanced challenges inherent in optimizing multi-agent systems for this task. These findings serve as an overview of current Reinforcement Learning (RL) optimization strategies, and could contribute valuable insights to the effective deployment of RL in complex, collaborative robotic scenarios.