This paper presents an effective neural adaptive approach for robot force control with changing/unknown robot-environment interaction dynamic properties. In this approach, a multilayered neural network controller is trained at first off line from data collected during contact motion in order to perform a smooth transition from free to contact motion. Then, an adaptive process is implemented online through a desired impedance reference model such that the closed-loop system maintains a good performance and compensates for uncertain/unknown dynamics of the robot-environment interaction. The effectiveness of the proposed approach has been evaluated for the force control of a 6 DOF (Degree Of Freedom) C5-links parallel robot executing rectangular peg-in-hole insertions with weak tolerances. The experimental results demonstrate that the robot’s skill improves effectively and force control performances are good even if robot-environment interaction dynamic properties change.