This paper presents a novel reactive motion planning framework for navigating robots in unknown and cluttered 2D workspace. Typical existing methods are developed by enforcing the robot staying in free regions represented by the locally extracted ellipse or polygon. Instead, we navigate the robot in free space with an alternate starshaped decomposition, which is calculated directly from real-time sensor data. Additionally, a roadmap is constructed incrementally to maintain the connectivity information of the starshaped regions. Compared to the roadmap built upon connected polygons or ellipses in the conventional approaches, the concave starshaped region is better suited to capture the natural distribution of sensor data, so that the perception information can be fully exploited for robot navigation. In this sense, conservative and myopic behaviors are avoided with the proposed approach, and intricate obstacle configurations can be suitably accommodated in unknown and cluttered environments. Then, we design a heuristic exploration algorithm on the roadmap to determine the frontier points of the starshaped regions, from which short-term goals are selected to attract the robot towards the goal configuration. It is noteworthy that, a recovery mechanism is developed on the roadmap that is triggered once a non-extendable short-term goal is reached. This mechanism renders it possible to deal with dead-end situations that can be typically encountered in unknown and cluttered environments. Furthermore, safe and smooth motion within the starshaped regions is generated by employing the Dynamical System Modulation (DSM) approach on the constructed roadmap. Through comprehensive evaluation in both simulations and real-world experiments, the proposed method outperforms the benchmark methods in terms of success rate and traveling time.