A major obstacle to the implementation of an intelligent, multi-sensor integrated robotic system is the lack of a uniform environment for developing the diverse system components and for efficiently communicating between these components. Our work is concerned with the solution through mapping such system onto a single parallel processor. We present a graphical programming environment, the Function Network Programming Environment, for developing and hierarchically organizing the computational modules of complex, asynchronous systems. In this environment, dependencies between modules are described graphically by interconnected function nodes; hence, parallelism is expressed naturally. Implementation on the BBN Butterfly parallel processor is also discussed, and a hypothetical development cycle for a multi-sensor integrated robot is described. It is expected that this environment will greatly facilitate the design and development of sophisticated, multi-sensor integrated robotic systems as well as enhance human comprehension of such systems.

The design and operation of an experimental robot workstation that adaptively responds to a disordered or changing environment are presented. The adaptive control is provided by the use of several sensor subsystems in a distributed architecture workcell. Included in the workstation are a commercial 6-axis robot, a microprocessor-based vision subsystem, an acoustic-ranging sensor subsystem, and a force-torque sensing subsystem; each subsystem is microprocessor based, and supervisory control is provided by a workcell host computer. Five demonstrations of adaptive control are described, including real-time path modification by single sensors and integrated use of multiple sensors.

A technique is described for improving the accuracy of the path followed by a robot arm in precision tracking of an object surface contour. Acoustic pulse echo data from a sensor mounted on an end effector was processed to compute distance to an object surface several inches from the sensor. Real-time path modification was then based on this computed distance. An eddy current proximity sensor positioned approximately 30 mils from the surface was used to monitor the acoustic sensor and robot performance. Both sensors produced identical results in tracking variations in distance from the surface with the robot in open loop mode for a linear path over a precision ground surface for which the robot was taught the two endpoints of the trajectory. Variation from the linear path was as large as 10 mils. With the acoustic range sensor providing input for a control algorithm to maintain a fixed offset as the robot traversed the surface, path deviation was reduced to +/- 3 mils using single measurements in the real-time mode. This is attributable to the inherent noise level of the sensor system. Averaging 10 measurements reduced the standard deviation in position measurement and control to +/- 1.2 mils.