Abstract: To increase automation in game design, this paper proposes a sampling-based motion-planning approach that works in conjunction with physics-based game engines. The approach automatically computes a sequence of motions that enables a virtual agent to reach a desired destination while avoiding collisions. The use of physics-based engines as the underlying simulator results in physically-realistic motions that take into account the motion dynamics, friction, gravity, and other forces interacting with the virtual agent. To account for the increased complexity and achieve computational efficiency, the approach expands a motion tree from the initial state to the goal using discrete abstractions as a guide in a best-first search fashion. Parametrized motion controllers are combined with randomized sampling to enable the motion planner to expand the motion tree along different directions. Comparisons to related work show significant computational speedups.