Mapless Navigation of Modular Mobile Robots using Deep Reinforcement Learning

The modular mobile robots can utilize their many degrees of freedom in variable configurations to adapt to unknown environments. The mapless navigation task implemented by modular mobile robots is more challenging than traditional mobile robots, since the motion planner requires more efforts on the coordination of modular units to guarantee integrity and flexibility. In this work, we address this challenge by proposing a novel mapless navigation framework which contains a two-level planner based on deep reinforcement learning (DRL). In our framework, the upper-level planner determines a high-level task mode and the lower-level outputs motion control commands, and the two planners update at different time scales. The results of simulated experiment conducted on Webots platform show that, the proposed framework allows modular mobile robots to use raw sensory information to efficiently explore and accomplish multitask in unknown environments.