In this paper we demonstrate a method for fine-grained modeling of a synthetic agent?s physical capabilities — running, jumping, sneaking, and other modes of movement. Using motion capture data acquired from human subjects, we extract a motion graph and construct a cost map for the space of agent actions. We show how a planner can incorporate this cost model into the planning process to select between equivalent goal-achieving plans. We explore the utility of our model in three different capacities: 1) modeling other agents in the environment; 2) representing heterogeneous agents with different physical capabilities; 3) modeling agent physical states (e.g., wounded or tired agents). This technique can be incorporated into applications where human-like, high-fidelity physical models are important to the agents? reasoning process, such as virtual training environments.