Task-Specific Uncertainty Maps and Bayesian Reinforcement Learning for Autonomous Navigation in Unknown Spaces

In dynamic and information-incomplete unknown spaces, enabling autonomous mobile robots to perform navigation tasks while balancing safety and efficiency remains a core challenge in the field of intelligent navigation. Existing reinforcement learning and graph search methods generally assume a homogeneous distribution of environmental uncertainty, overlooking the varying requirements for localization accuracy and risk tolerance across different task regions, and lack reliable quantification of model and perception errors, making it difficult to handle complex scenarios. To address this issue, this paper proposes a navigation framework that integrates task-specific uncertainty maps with Bayesian reinforcement learning. First, a task-oriented spatial uncertainty map is constructed, assigning acceptable prediction variances to each location based on a comprehensive evaluation of environmental risks, complexity, and constraints. Second, Bayesian neural networks are employed to perform probabilistic modeling of robot dynamics and environmental dynamics, calculating the mean and variance of state transitions in real-time through variational inference. Then, a risk-sensitive value function is defined in the Bayesian belief space, incorporating a penalty term to enforce cautious actions in high-variance regions, thereby designing a decision-making algorithm that integrates task requirements, model uncertainty, and reinforcement learning. Simulation and experimental results demonstrate that the proposed method significantly improves navigation success rates, reduces average path lengths, and lowers collision risks in unknown environments. Compared to seven baseline methods, including DQN, PPO, SAC, CVaR RL, conservative policy optimization, and TSUM-Guide RL, the proposed approach achieves an average success rate improvement of over 10\\% and a path length reduction of 8\\%, showcasing strong innovation and practical value.