Online Monte Carlo Planning for Autonomous Robots: Exploiting Prior Knowledge on Task Similarities

Planning in large state spaces is a key problem in robot autonomy applications. In this paper we evaluate an extended version of the Partially Observable Monte Carlo Planning (POMCP) algorithm on simulated (Gazebo) and realenvironments for instances of Rocksample, where a TurtleBot is used as an agent. The extended POMCP planner exploits prior knowledge about task similarities to reduce the explored state space improving robot performance. Results show that the proposed method significantly outperforms the standard POMCP with an improvement of average discounted return up to 60.7%. This improvement implies reduced number of steps performed by the robot, shorter path lengths, reduced total running times and better energy management in long-term deployments. The main contributions are the integration of the extended POMCP planner into simulated and real robotic platforms, and performance comparison between standardand extended POMCP planners in these environments.