A computationally efficient method to achieve reflexive collision avoidance is described. The system consists of four parallel, asynchronous control layers: a conventional servo tracking controller, a reactive trapezoidal velocity profiler, a joint-by-joint reflexive collision avoidance layer, and a guiding potential-function layer. The lowest two layers guarantee that the robot will not overshoot any joint-space position setpoint. The higher two layers compute successive set points based on a continuous, online inspection of configuration space. The guiding potential-function layer incorporates task-space-based attraction to a hand goal, while the reflexes permit use of kinematic redundancy for obstacle avoidance. In many cases, no explicit path planning is required to achieve competent, collision-free motion to a task-space goal while moving in a cluttered environment. Experimental results are described for the physical implementation of the system on the first four joints of a kinematically redundant robot.>