Autonomous Neural Network Controller for Adaptive Material Handling
Abstract
Current methods in motor control have problems dealing effectively with highly variable dynamic inertial interactions between multijointed robots and payloads. We are developing an autonomous neural network controller that can overcome these difficulties by learning to anticipate the inertial interactions from its own experience. The neural network controller win allow robots to handle diverse payloads in uncertain environments to benefit a wide variety of material handling applications. Our target application is bin-picking, the grasping of an object from a bin containing many randomly oriented objects and placing it at a desired location. During this quarter of the SBIR Phase II contract, we focused on the dynamic control aspect of the problem by extending our working implementation of the neural network controller from the Phase I effort. Using a commercially available scara-type robot, we demonstrated a functional prototype of the neural network controller for realizing point-to- point control. The controller design consists of dynamic position and velocity servos in parallel with an adaptive neural network controller for each joint. The controller adaptively learns to compensate for the dynamic inertial interactions with different payloads through its own experience. Using two joints of the scara robot, the controller achieved a position accuracy of 0.2% of the joint range, a timing accuracy of within 5% of the requested movement time, and an end-point stability of within 8% of the maximum planned velocity. This performance was measured on both joints after only 150 training iterations with a movement that had large dynamic coupling forces between the scara links.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 28, 1992
- Accession Number
- ADA248944
Entities
People
- James Kottas
- Michael Kuperstein