Augmenting a Nominal Assembly Motion Plan with a Compliant Behavior

Gordon A. Dakin, Robin J. PopplestoneAAAI 1991

Abstract

This paper introduces a practical framework for robotic assembly that bridges the gap between idealized geometric motion plans and the uncertainties inherent in physical environments. By employing a two-phase approach, the methodology first generates a nominal geometric path and subsequently augments it with a fine-motion strategy. This strategy utilizes linearization of configuration space surfaces and linear programming to efficiently account for manufacturing tolerances and sensor errors, transforming computationally prohibitive problems into manageable verification tasks.

Introduction

In robotic assembly, the transition from theoretical models to factory floor implementation is often hindered by real-world imperfections. While a robot can follow a mathematically precise path in a zero-clearance model, actual parts possess manufacturing tolerances, and robotic controllers exhibit inherent errors. Traditional approaches, such as the Lozano-Pérez, Mason, and Taylor (LMT) methods, provide elegant theoretical solutions but suffer from extreme computational complexity ($O(2^{2^n})$), rendering them impractical for complex tasks.

This work proposes a more efficient 2-phase approach. The first phase involves generating a "nominal" trajectory based on pure geometry. The second phase refines this trajectory into a "fine-motion strategy" capable of handling uncertainty without the exponential overhead of previous methods.

Methodology

The core of the proposed framework is the representation and manipulation of the Configuration Space (C-space).

Linearization of C-Surfaces

The authors observe that because assembly clearances are typically very small, the complex, curved boundaries of contact states can be effectively approximated as flat hyperplanes. This linearization is the fundamental driver of the system's speed and efficiency.

Generalized C-Space and Linear Programming

To account for uncertainty, the C-space is expanded into a "Generalized C-Space" where model errors (e.g., a part tolerance of $\pm 0.1mm$) are treated as additional dimensions. By formulating the geometry as a set of linear equations, standard Linear Programming (LP) tools can be used to:

  1. Verify the physical possibility of specific contact sets (e.g., vertex-to-face).
  2. Identify adjacent contact states to facilitate plan transitions.
  3. Ensure commanded velocities are "safe" by checking if the "forward-projection"—the volume of space a part might occupy given error bounds—intersects any forbidden obstacles.
Figure 4: A nominal assembly configuration showing tight clearances between polyhedral parts.

Key Insights: Uncertainty as a Volume

A pivotal insight of this research is that tight fits, while challenging for execution, simplify the underlying mathematics by justifying linear approximations. The space of valid, non-overlapping positions (the "LEGAL" set) is partitioned into a union of convex volumes. This allows the robot's path to be treated not as a line, but as a "swept volume" or polytope. If this volume remains clear of "bad" contact states, the motion is guaranteed to be successful despite sensor or manufacturing errors.

Figure 5: Examples of different contact states that the fine-motion strategy must navigate during the assembly process.

Results and Implementation

The methodology was implemented in the POPLOG environment on a Sun 3 workstation using geometric models of polyhedral parts. Key results include:

  • Computational Efficiency: Replacing exponential-time algorithms with linear programming significantly improved scalability for complex polyhedral parts.
  • Verification Capabilities: The system functions as a robust tool for "proving" the success of fine-motion plans. If the linear programmer finds no intersection between the robot's potential error-filled path and an obstacle, the plan is verified.
  • PNT Generation: The authors demonstrated the generation of Perturbed Nominal Trajectories (PNT), which guide robots through sequences of contact states (like sliding a part along a wall) to reach a goal.

Conclusion and Future Work

While highly effective for polyhedral parts, the current framework is limited to flat faces and straight edges. Future research aims to extend these linearization techniques to quadric surfaces (e.g., cylinders and spheres) common in mechanical engineering. Additionally, while the tool excels at verification, the initial selection of command velocities still relies on human intuition, suggesting a need for automated heuristic generation in future iterations.

Summary updated

BibTeX
@inproceedings{dakin1991augmenting,
  author = {Gordon A. Dakin and Robin J. Popplestone},
  title = {{Augmenting a Nominal Assembly Motion Plan with a Compliant Behavior}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1991},
}
Figure