Uncalibrated visual servoing for full motion dextrous robot systems with tracking cameras

Visual servoing methods are reviewed along with their limitations in low-cost uncalibrated situations. A description of the camera model is then given including an overview of visual Jacobian calculations. The problem of selecting suitable gains for implementation in real-time is developed. The model system used is of the anthropomorphic service robot HERMES. It is highly non-linear and redundant with 6 d.o.f. in the camera system and 6 d.o.f. in each of its two arms. The selection of gains is difficult due to the non-linearity of the Jacobian over the complete robot. A conservative controller for the camera is developed and a single measure for gripper orientation is outlined along with its limitations. Two solutions to the problems of alignment in visual servoing over a large workspace are given. The first weights the alignment heavily during the initial stages and the second uses a starting point grid derived by „visually voxellating“ the workspace. Results of exhaustive and local motion tests are then given. The simple variable step size gain is found to be effective. The proposed alignment method improved the visual servoing alignment performance.