Quantization effects can reduce the achievable stiffness of virtual walls. Phase delays introduced by filtering are also an issue. This paper presents a non-linear model-based observer that produces smooth position and velocity estimates with very little delay based on measured position and force signals. Compensation for Coulomb friction and motor saturation is incorporated into the estimator. Use of the estimator in the implementation of a virtual wall yielded higher wall stiffnesses and better performance. Model-based estimator design was possible due to the design of the manipulator. The three degree-of-freedom manipulator employed is direct-drive, gravity-balanced, and dynamically-decoupled with nearly linear dynamics. The robot structure itself is employed as a force sensor, reducing the cost of the device.