Semantic compressive sensing for robot motion control

Abstract

Robots are changing our lives in many areas, such as manufacturing industry, physical surgery, military operations, and logistics. In a robot motion control system, feedback, especially visual information, plays an important role in improving system performance by controlling robotic motion through a closed-loop strategy. However, two major problems exist. First, feedback images are basically sets, not even smooth. Traditional visual feedback control approaches mainly use vectors instead of sets to represent the feedback images. Assigning vectorial characteristics to images enables us to employ the control theories established on differential equations to analyze and control a system. However, it requires to extract visual features, which is often time-consuming and may bring extra extraction errors when image features are not explicit. On the other hand, feedback data often have large size, which decreases the feedback rate, thus reducing system real-time performance. This influence is obvious for embedded systems, such as unmanned aerial vehicles whose computational power is limited.

To overcome the first problem, a feedback control strategy is proposed in the space of sets, where feedback images are represented by sets instead of vectors. The commonly used differential equation is not suitable to analyze and design control laws for systems in the space of sets because the vector space linear structure does not exist. Therefore, the mathematical tool, mutational analysis, is used to describe the systems. The techniques of stability analysis and controller design are proposed within the framework of mutational analysis. Then the developed control strategy is applied to visual feedback control problem by representing feedback images directly with sets to avoid the extraction of visual features.

To cope with the second problem, a novel compressive sensing method, which can sample and compress feedback signals simultaneously is proposed to obtain compressed feedback signals. This sensing approach can reduce the computational load of a control system and thus improve the real-time performance of the system by taking the prior knowledge of feedback data into account. It also expends the applications of compressive sensing to non-sparse signals by eliminating the sparsity constraint of traditional compressive sensing approaches.

Finally, based on the developed control strategy and sensing method, a compressive feedback control framework in the space of sets is proposed to control the motion of a robot. In this framework, the proposed compressive sensing approach is used to obtain the compressed signals, which are used directly as the feedback for the set-based robot motion control strategy to reduce the computational load and increase the feedback rate. The stability property of the system with compressive feedback data is analyzed. As an application, the problem of visual control with compressed feedback is addressed within this framework.