Comparison of image-based and position-based robot visual servoing methods and improvements

Vision-based robot control, or visual servoing, significantly increases flexibility of conventional robot applications. Two major methods, image-based visual servoing (IBVS) and position-based visual servoing (PBVS), have been co-existed for a long time without a systematic and complete comparison. This work first contributes by proposing a common comparison framework based on the generic sensory-task-space robot control approach. A complete comparison between IBVS and PBVS is investigated on this common basis for the first time. The simulation and experimental results verify, that the two methods are quite comparable in system dynamic stability, robustness, and sensitivity; however, there exist significant differences in the Cartesian and image space dynamic performance. Two common characteristics that affect both methods in a similar manner are identified. First, both methods have difficulty in handling large motion commands. Second, accurate a prior knowledge of the target model is necessary for proper IBVS and PBVS control.; To deal with the above problems, the next major contributions of this work are the novel hybrid motion control and planning strategies and the online unknown target model estimation strategy. The hybrid motion control strategy allows direct avoidance of image singularities and image local minima during large motions. A complete set of Cartesian, image, and robot joint constraints under a complex visual servoing scenario are considered in the hybrid motion planning strategy. IBVS or PBVS can be applied to track the planned trajectories for complete assurance of all constraints in the presence of large motion commands. The decoupled extended Kalman filter-based unknown target model estimation strategy allows estimation of the target model and relative pose simultaneously under continuous robot dynamic motion. The method is robust to large initial estimation errors and provides consistent and accurate target model estimation for optimal pose estimation as required in PBVS. The simulation and experimental results demonstrate the usefulness and the effectiveness of these improvements to traditional IBVS and PBVS methods.