The main factors that attribute to the accuracy of a direct-drive robo
t are the geometry, tumbling joint motion, and encoder errors. This wo
rk presents a model that comprises a description of these factors conc
entrating on tumbling and encoder errors. The model is based on a new
systematic approach to model the geometry. The model is non-redundant
and free of singularities. Using the measurements of a single robot ax
is all the parameters of the model are identified. Due to the high rep
eatability the confidence interval of the found parameters is small. A
fter calibration the improved accuracy is very close to the robot's re
peatability.