Intelligent Automation, Inc. (IAI) presents a system that learns from demonstrations of a human operator and then produces a sensor-driven program that can control a robot without operator intervention. The system generates a procedural program, identical in function to a procedural program developed in a conventional robot programming language, that can be easily re-used in similar but different application tasks. In essence, based on the skills demonstrated by a human teacher, a robot can be automatically programmed to perform complex assembly tasks. Robot users with no programming skills will be able to "bootstrap" themselves to a semi-autonomous operation after a small number of taught examples. In addition, as the task requirements change, the robot can be re-programmed as easily as it was initially programmed. Unlike the conventional program development process, the operator need not postulate about the value of sensor feedback variables required at a particular level of the controller. In fact, the operator need not be familiar with the syntax and semantics of the programming language at all.

This project develops a "teach-by showing" approach that permits unskilled users to program robots by physically demonstrating examples of a task to a trainable controller. Each example may be shown in one of three ways: by direct demonstration, by placing the robot in a zero-force mode and physically leading it through the task, or by operating the robot in a master/slave mode. During each demonstration, both the sensor inputs to the robot controller and the control outputs from the controller are presented to the training system. Following the demonstrations, the training system then produces a prototype procedural program suitable for downloading into a commercial robot controller. Unlike the conventional program development process, the operator need not postulate about the value of sensor feedback variables required at a particular level of the controller. Examples of successful, but perhaps quite different task strategies can be demonstrated by the human teacher until the robot system is capable of successful autonomous operation.

An algorithm through which above-knee (AK) amputees may exercise voluntary and subconscious control of knee flexion and extension with their prosthetic device is presented in this paper. It is hypothesized that spatial patterns of electromyograms (EMG) from muscles above the site of amputation can be separated into groups representing 1) knee flexion, 2) knee extension, and 3) hip action involving no knee activity. An experiment was designed and an apparatus constructed to apply known torques in various configurations to the right leg of normal subjects while sampling EMG from selected muscle groups. Various pattern classification techniques were then applied to judge the degree of separability among each of the three classes of patterns. It was found that the classification can be performed virtually without error depending upon the complexity of the training procedures used to develop the classifier parameters. In addition the magnitude of the classifier response (in the positive sense only) is correlated with the applied knee torque in the direction for which the classifier is designed. Similar results have been found with an above-knee amputee as a subject.