From Kinematics to Shape: An Approach to Innovative Design

Leo Joskowicz, Sanjaya AddankiAAAI 1988

Introduction

Presented at AAAI-88, this research represents a foundational moment in automated mechanical design. It moves beyond simple parameter optimization to demonstrate how computers can "invent" new shapes to achieve specific mechanical behaviors. In the late 1980s, Computer-Aided Design (CAD) was largely a passive tool. While "routine design" systems could optimize existing parts, they lacked the capability to generate new features, such as adding a notch to a wheel to enable locking. This paper addresses the "kinematic gap": translating functional requirements into specific physical geometries.

Methodology: Configuration Space as a Bridge

The core of the approach is the use of Configuration Space (C-space). In a C-space, every point represents a possible placement of two objects. Regions are categorized as "free" (no contact) or "boundary" (objects in contact).

The transformation of Configuration Space (C-space) from an initial state to a modified state that satisfies new kinematic constraints.

The design process follows a structured algorithm:

  1. C-Space Comparison: The system identifies differences between the "Actual C-space" (current shapes) and the "Desired C-space" (target behavior).
  2. Backtracking Search: A search process modifies object contours by adding or deleting line segments and circular arcs.
  3. Elementary Contact Mapping: A lookup table suggests specific geometric features (e.g., vertex-arc or edge-arc contacts) to produce desired C-space boundaries.
  4. Qualitative Reasoning: The method handles vague instructions (e.g., "Object B should move up when Object A rotates clockwise") by translating them into C-space regions.

Evaluation and Results

The authors evaluated their method using synthetic design scenarios, most notably the Disc and Rectangle problem. In this scenario, a rotating disc must be modified to lock a translating rectangle at specific orientations ($0$ and $\pi/2$).

Initial and modified shapes of a disc (A) and rectangle (B), showing the automated addition of slots to satisfy locking requirements.

The algorithm successfully "invented" the necessary slots in the disc, a task previously requiring human intuition. Key findings include:

  • Automated Feature Discovery: The system can change the topology of a part (adding holes or edges) rather than just adjusting dimensions.
  • Computational Efficiency: For common cases like convex objects, the algorithm runs in linear or polynomial time.
  • Causal Mapping: Abstract state diagrams can be mathematically mapped to physical C-space regions, linking common-sense physics with geometry.

Discussion and Future Work

While groundbreaking, the method is limited to 2D objects with a single degree of freedom and contours composed of straight lines and circular arcs. The general backtracking algorithm can also face exponential complexity as features increase. Future directions include extending the framework to 3D geometry and complex kinematic chains with multiple moving parts. By treating C-space as an intermediate representation, this work strips away geometric complexity and frames mechanical design as the "shaping" of interaction space.

Summary updated

BibTeX
@inproceedings{joskowicz1988from,
  author = {Leo Joskowicz and Sanjaya Addanki},
  title = {{From Kinematics to Shape: An Approach to Innovative Design}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1988},
}
Figure