We present a principal component-based method for generating, in real time, fast robot motions that minimize power consumption. Given a dynamic model of a robot, a sufficiently large set of torque-minimum motions are first obtained for preselected initial and final positions that also achieve minimum time while avoiding actuator saturation. These motions are then clustered according to the trajectory endpoints and shape. A principal component analysis is performed for each motion cluster, and the dominant principal components are used as basis functions in a linear interpolation scheme for generating fast, torque-efficient motions between arbitrary initial and final positions. Results obtained for both a six-axis industrial manipulator and a wheeled mobile robot demonstrate that nearly optimal motions can be obtained in real-time using this scheme.