Automated guided vehicles (AGVs) are a primary focus of the robotics industry. These vehicles have previously lacked vision-based guidance and depended on structured environments or changes to existing plant infrastructure to function. For an automated forklift to pick up a pallet of materials, for example, special jigs and attachments are required. But automating forklift and tug operation in existing industrial plants would decrease labor and vehicle maintenance costs, allow for more efficient plant operation through more rapid and reliable delivery of materials, and help support round-the-clock production.
At the National Robotics Engineering Consortium (NREC; Pittsburgh, PA), such a program is underway. Dubbed the Automated Material Transport System (AMTS), the project aims to develop an automatic guided vehicle that will work in less structured environments than existing AGVs and require little or no supporting plant infrastructure. Teaming the Ford Motor Co. (Dearborn, MI), Hyster (Memphis, TN), FMC (Chicago, IL), NASA Jet Propulsion Laboratory (Pasadena, CA), and NASA Ames Research Center (Moffett Field, CA), the AMTS program is integrating computer-vision technology and planning software, a vehicle-position estimation system, a pallet acquisition-and-stacking system, and application-specific software.
Vision technology for position estimation aims at allowing vehicles to position themselves in a large industrial environment by creating an image of the floor area called a map. During operation, the AMTS vehicle acquires images of the floor and locates these images in the on-board map. Floor mapping is neither difficult nor expensive, requires no changes to the plant's infrastructure, and costs significantly less than other navigation alternatives.
Says Alonzo Kelly of the Robotics Institute of Carnegie Mellon University (Pittsburgh, PA), one of the principal developers of machine-vision software for the project, "Even a decade ago, it may have sounded absurd to propose navigating from a 10-Gbyte image of a factory floor. However, we have now reached the point where memory, processing, and imaging hardware are both adequate and inexpensive enough to do so.
"Today, a double-sided, double-layer DVD-18 DVD ROM disk can store 17 Gbytes, and the latest Pentium MMX-based computers can perform template correlation in a single instruction," explains Kelly. In the software developed by Kelly, a technique called mosaic-based localization is used to position the robot. In this technique, a mosaic of a large area of the factory floor is created and stored in global memory.
According to Kelly, 10 km of guidepath at 1-mm resolution can be stored in 10 Gbytes of memory. To match the acquired image or template with the mosaic region, normalized crosscorrelation is computed between the images. The autocorrelation score image is generated from the difference of the highest and the second-highest autocorrelation peak in a 17 x 17 pixel search window. This information is used to position the vehicle on the factory floor.
In 1998, a demonstration of automated trailer loading, unloading, and stacking was completed. Last year, the system was upgraded to include obstacle avoidance and multi-vehicle coordination. This year, the AMTS program will include a tug AGV pilot program at an automotive-assembly plant. Next year, a forked AGV pilot program is planned for this plant.
