Responsive and sensitive force control is essential for sophisticated robot tasks such as assembly and fine manipulation. However, this kind of control is very difficult for manipulators driven by electric motors with geared transmissions, primarily due to friction in the drive train. Yet these types of mechanical systems are often the most practical for many applications. Two primary types of control strategies have been used to date for force control with these kinds of systems. One uses the force error to adjust the torque of the drive motor. The other utilizes force feedback to command motor positions, relying on compliance in the system for translating position to force. The former approach typically results in sluggish, noisy control and the latter approach typically exhibits poor bandwidth. Neither approach adequately addresses manipulators which are moderately back-drivable. We propose an adaptive force control system based on the "backpropagated adaptive critic" architecture, which utilizes principles from dynamic programming and neural networks to achieve real-time adaptation. This approach requires few assumptions about plant dynamics, and a wide variety of sensor inputs can easily be incorporated. Furthermore, effects which depend on past as well as current states, such as backlash, can also be accommodated. Higher performance force control for common electric manipulators (and hands/ grippers) can significantly enhance the use of these devices in automation and teleoperation. The general adaptive control approach could also be applied to many difficult control problems in the government and industry.
Keywords: Force Control, Compliance, Adaptive Critics, Intelligent Control
