Off-Policy Evaluation With Online Adaptation for Robot Exploration in Challenging Environments

Autonomous exploration has many important applications. However, classic information gain-based or frontier-based only relies on the robot current state to determine immediate goal, which lacks capability of predicting value future states and thus leads inefficient decisions. This letter presents a method learn how “good” are, measured by function, provide guidance for in real-world challenging environments. We formulate our work as an off-policy evaluation (OPE) problem (OPERE). It consists offline Monte-Carlo training data performs Temporal Difference (TD) online adaptation optimize trained estimator. also design intrinsic reward function based sensor coverage enable gain more with sparse extrinsic rewards. Results show that enables predict so better guide exploration. The proposed algorithm achieves prediction performance compared state-of-the-arts. To best knowledge, this first time demonstrates dataset subterranean urban