Shape memory alloys (SMA) are a class of smart material having the unique ability to return to a predefined shape when heated. SMA based actuators have the potential to be very compact and low weight. As a result, much research has been devoted to the design of SMA based actuators; however, commercialization has been largely impeded by the small strain capacity inherent to SMA. To address this deficiency, this paper conveys the design of a large-strain SMA actuator (in excess of 30%) whose feasibility is investigated by integrating the actuators as artificial muscles in a two link anthropomorphic arm. The ensuing experimental results indicate that the actuators show great potential for a variety of emerging applications. Finally the design of an SMA based dextrous robotic hand evaluation facility is proposed, and provides a case study illustrating how smart structures provide a superior alternative to conventionally voluminous and heavy prosthetic actuators.
In robot-manipulator teleoperation, vision-based tracking of the human operator motion offers a non-contacting approach that permits unhindered operator motion. To control the robot manipulator, the three-dimensional (3D) position and orientation of the arm of the operator is required. This paper presents a neural-network (NN) based method of determining the orientation of the human hand using non-invasive markerless vision-based tracking. The tracking method uses images of the hand from two fixed cameras to determine three angles of hand orientation. The neural network processing to determine the hand orientation consists of five procedures. First, a preprocessing system performs basic transformations on the input images to prepare them to be interpreted by the neural network. Secondly, an unsupervised neural network extracts relevant local features necessary to recognize the input patterns. Thirdly, a self-organizing neural network combines the local features of the previous network to identify the global pattern. Next, a modified radial-basis function (RBF) neural network calculates the probabilities that a given input pattern corresponds to each basic pattern, for which the RBF NN was trained. Finally, the orientation of the hand is interpolated between these basic patterns by calculating the weighted average of the most probable configurations identified by the RBF NN.