Introduction
Multi-Agent Reinforcement Learning (MARL) addresses the decision-making for multiple agents in a common environment moving collectively [1, 2]. Many MARL algorithms were developed in recent years [3, 4] while only few were tested on real robots [5, 6]. In general, training Reinforcement Learning (RL) policies on real robots is a tedious and time consuming task [7]. In addition, the robot must work for a very long time which may cause damage and wear, and may pose danger. The problem is even more challenging in a multi-agent system where individual data is required from each agent. Training in simulation, on the other hand, is a compelling solution where the data is acquired at a lower cost [8]. Simulation-based learning provides a cost-effective way to collect data through interactions with the environment. Such an approach, for instance, was used for obtaining control for an autonomous vehicle using a simulation with synthetic images [9]. A similar approach was used for autonomous soil excavation [10]. However, simulations rarely capture reality and the trained policies are usually poorly transferred [11]. This problem is even worse for open-source hardware such as for underactuated robotic hands [12] and mobile robots [13]. Such hardware is usually 3D printed accompanied by low-cost actuators which impose many fabrication uncertainties in, for example, friction, size, mass and compliance. Therefore, acquiring analytical models for these systems is a challenging problem, even for experienced practitioners, leading to the lack of a good simulator [14].
to read the full article- Enter Here