Abstract: Robot path planning in unknown and dynamic environments is frequently challenged by perceptual uncertainty and stochastic obstacles, which significantly constrain both navigation robustness and real-time responsiveness. To address these issues, this paper proposes a path planning framework for partially observable environments that integrates belief entropy–based risk quantification with an improved A* algorithm. First, the robot navigation problem is formulated as a partially observable Markov decision process (POMDP) that enables the robot to represent environmental uncertainty through probabilistic belief states. By employing Bayesian probabilistic inference, the system continuously updates its belief about the surrounding environment based on partial observations to maintain a probabilistic representation of obstacle distribution and reduce the impact of incomplete sensing information. This probabilistic representation allows the robot to reason about potential environmental risks and make more informed navigation decisions even when the available sensory information is incomplete or partially unreliable. On this basis, the traditional A* evaluation function is reconstructed by introducing belief entropy as a dynamic risk compensation term. This modification allows the algorithm to incorporate environmental uncertainty directly into the heuristic search process and enables the robot to jointly optimize path cost and safety gain during navigation. Compared with conventional heuristic search strategies that only consider the geometric distance or traversal cost, the proposed method introduces an uncertainty-aware planning mechanism that discourages the robot from passing through regions with high environmental ambiguity. In addition, third-order Bézier curves are used to smooth the generated paths, thereby improving trajectory continuity and making the planned routes more suitable for physical robotic platforms that require smooth motion control and kinematic feasibility. A k-dimensional tree spatial indexing structure is adopted to accelerate neighborhood searches during collision detection and obstacle querying to effectively reduce the computational complexity from O(n) to O(log n). Extensive simulation experiments are conducted to evaluate the performance of the proposed framework. Results show that in grid-based environments, the proposed algorithm achieves an average single-step decision time of approximately 0.01 s, which is more than two orders of magnitude faster than the classical POMCP algorithm. In scalability stress tests where the map area is expanded by nine times, the peak memory consumption remains stable at approximately 0.12 MB, demonstrating excellent scalability and suitability for resource-constrained embedded systems. Robustness analysis further indicates that even in high-noise environments with elevated sensor failure rates, the algorithm can still maintain a task success rate of approximately 65%, with a performance degradation slope of only −0.54, showing strong tolerance to perception uncertainty. Visualization results confirm that the proposed framework effectively balances obstacle avoidance safety and goal-directed convergence, allowing the robot to avoid uncertain regions while maintaining efficient progress toward the target. Overall, by integrating probabilistic belief modeling with efficient heuristic search, the proposed method provides a practical and computationally efficient solution for real-time robust navigation of unmanned robotic systems operating in uncertain and partially observable environments. It is particularly suitable for complex industrial and service robotics scenarios with significant uncertainty, dynamic disturbances, and limited onboard computational resources in current practical applications.