Abstract:

In practical applications, mobile robots must be able to navigate and explore complex surroundings autonomously. Prior approaches to addressing autonomous exploration involved having the robots keep an internal map of the environment, which it would subsequently navigate using localization and planning techniques. However, these methods frequently make several assumptions, require a lot of work, and do not consider failures. Contrarily, learning-based approaches enhance as the robot interacts with its environment. Still, they take longer to converge to the best solution, have significant communication and update costs, and are therefore challenging to implement in practical settings due to their considerable sample complexity. We propose CQLite, a light version of cooperative Q-learning over a coverage-weighted reward function with limited communication overhead, to address the need for fast convergence. We analyze the approach on empirical grounds to guarantee full coverage, fast convergence, and reduced computational complexity. We also validate our approach with simulation and real-world indoor map exploration using multiple robots and compared our approach with commonly used Rapidly-exploring Randomized Trees (RRT) and state-of-the-art Deep Reinforcement Learning (DRL) techniques for multi-robot map exploration.

Codes and Dataset: Github Link