Introduction
In the early 1970s, industrial automation was characterized by rigid, "microscopic" programming. Robots were typically programmed with precise, pre-defined coordinates for every movement, a method suitable for mass production but inefficient for small-batch manufacturing where product designs frequently changed. This 1971 paper from Hitachi’s Central Research Laboratory describes "HIVIP," a landmark system designed to bridge the gap between abstract human instructions—specifically engineering drawings—and physical execution. HIVIP represents a shift toward "macroscopic" goal-oriented instruction, where a robot is given a target state and autonomously determines the necessary steps to achieve it.
System Architecture
The HIVIP system integrates three primary subsystems: the EYE (visual system), the BRAIN (processing unit), and the HAND (manipulator).
The hardware configuration includes:
- EYE: Two vidicon cameras. Camera 1 is dedicated to reading 2D engineering drawings, while Camera 2 monitors the workspace.
- BRAIN: A HITAC 7250 digital computer with 32k words of core memory.
- HAND: A 7-degree-of-freedom (DOF) articulated arm equipped with a parallel-jaw gripper.
Methods and Cognition
The system's operational flow involves four distinct stages:
- Instruction Recognition: The system analyzes a "three-view plan" (top, front, and side views) captured by Camera 1. It reconstructs a 3D internal model by resolving 2D line segments into a spatial representation. A significant challenge in this phase was the identification of "false points"—intersections in the 2D projection that do not correspond to actual physical vertices in 3D space.
- Object Recognition: Using Camera 2, the robot scans the assembly table to identify available polyhedral parts. It employs grayscale differentiation and a "trenching" method for edge detection, matching the physical blocks' dimensions and shapes to the components identified in the drawing.
- Decision-Making: To plan the assembly, the robot utilizes a "disintegration analysis" algorithm. Rather than planning the assembly from scratch, it calculates how the final 3D structure would be disassembled (creating a "disintegration tree"). Reversing this sequence provides a logical and collision-free assembly plan.
- Manipulation: The computer calculates the necessary joint displacements and approach vectors for the 7-DOF arm, executing movements in a continuous-path mode to assemble the parts.
Experimental Results
The HIVIP system successfully demonstrated the autonomous assembly of structures composed of white polyhedral blocks (prisms and grooved blocks). A typical assembly cycle required approximately 8 minutes, distributed as follows:
- Image Processing: 240 seconds
- Drawing Recognition: 20 seconds
- Object Recognition: 50 seconds per object
- Decision-Making: 10 seconds
- Assembly Movements: 180 seconds (average)
While slow by modern standards, this was a pioneering achievement given the hardware constraints of the era.
Discussion and Limitations
The research highlighted several key concepts that remain relevant in robotics, such as the distinction between macroscopic and microscopic instructions and the use of "visibility control" (an early form of visual attention). However, the system had notable limitations:
- Geometric Constraints: It was restricted to simple polyhedra with flat surfaces and straight edges.
- Environmental Sensitivity: The vision system required high-contrast lighting and simple backgrounds, lacking the robustness needed for real-world factory environments.
- Processing Speed: The authors identified that the overhead of FORTRAN and the sequential nature of the hardware limited industrial viability, suggesting a future need for parallel processing.
Despite these constraints, HIVIP proved that a robot could "imagine" a 3D object from a 2D drawing and autonomously execute a complex physical task, laying the groundwork for modern intelligent automation.
Summary updated