Geometric Reasoning and Organized Optimization for Automated Process Planning

Yasuyuki Maeda, Katsuya ShinoharaAAAI 1988

Abstract

Automated process planning serves as the critical link between digital design and physical manufacturing. This paper presents the ESPER (Expert System for Production EngineeRs) framework, which addresses the challenges of translating complex 3D geometries into machine instructions. By utilizing geometric reasoning to convert CAD models into symbolic networks and employing an organized optimization strategy with human-in-the-loop capabilities, ESPER bridges the gap between high-level design and factory-floor execution.

Introduction

In the evolution of Computer-Integrated Manufacturing (CIM), the transition from Computer-Aided Design (CAD) to Numerical Control (NC) tape generation has historically been a bottleneck. Two primary obstacles hindered automation: the "Geometry Gap," where standard AI logic struggled to interpret physical 3D shapes, and the "Optimization Explosion," where the search space for efficient machining paths became computationally intractable.

The ESPER Framework

The ESPER system introduces a dual-pronged approach to overcome these hurdles: geometric reasoning for feature recognition and a modular optimization strategy.

Geometric Reasoning and Feature Recognition

To enable symbolic reasoning over physical space, ESPER transforms CAD models through a multi-stage process. Complex 2.5D contours are first approximated and then decomposed into primitive geometric shapes, such as rectangles, using a greedy algorithm. These primitives are then organized into a relational network that captures adjacency and inclusion.

The multi-stage geometric reasoning process: (a) original contour, (b) feature approximation, (c) decomposition into primitive rectangles, and (d) relational network generation.

This network representation allows the system to apply heuristic rules to determine machining sequences—for instance, identifying whether a deeper pocket can be machined simultaneously with an adjacent shallower area.

Organized Optimization via Task Modules

ESPER partitions the planning process into 13 specialized task modules, including tool selection, milling order, and cutting path generation. Each module follows a two-stage execution:

  1. Candidate Generation: Heuristic rules are applied to prune the search space, identifying only the most viable machining options.
  2. Optimization: Specialized algorithms select the optimal candidate based on efficiency metrics, such as minimized machining time.

Human-in-the-Loop Interaction

Recognizing the limitations of purely automated systems, ESPER incorporates an interactive "User Action" system. This allows human engineers to intervene at specific breakpoints, modify tool selections, or adjust suggested paths. The system then dynamically re-calculates the remaining plan to maintain consistency with the expert's input, combining AI efficiency with human intuition.

Results and Discussion

The system was implemented using UTILISP and EXCORE and evaluated on 2.5D mechanical components made of aluminum and steel. ESPER demonstrated a robust capability to handle complex parts that were previously beyond the reach of automated planning, particularly in managing the transition from roughing to finishing operations. By pruning the search space before optimization, the system effectively avoided combinatorial explosions while maintaining high plan quality.

Conclusion

The ESPER system successfully demonstrates that geometric reasoning, when coupled with organized optimization and human interaction, can effectively automate the complex task of process planning. Future work will focus on extending these techniques to full 3D curved surfaces and developing more sophisticated graphical interfaces to visualize the planning logic in real-time.

Summary updated

BibTeX
@inproceedings{maeda1988geometric,
  author = {Yasuyuki Maeda and Katsuya Shinohara},
  title = {{Geometric Reasoning and Organized Optimization for Automated Process Planning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1988},
}
Figure