The interest in underactuated mechanisms has seen an increase in the robotics community in the application of agile dynamic motions such as human-like walking. Rapidly-exploring Random Trees (RRT) algorithms serve as a suc- cessful and practical approach for path planning in high-dimensional systems, but their applicability for systems with underactuation has been less success- ful. In this paper, we propose two modifications to the RRT algorithm which facilitate the search of a start-to-target actuation command policy for under- actuated systems. The first modification addresses the distance cost function which defines proximity of states. Specifically, we estimate the reachability property from one state to another with a simple geometrical heuristic. The second modification was geared to bias the distribution of the random samples in the state space towards the desired target states. We present the results on a torque-limited simple pendulum as well as on an underactuated model of a double pendulum.