Naive Kinematics: One Aspect of Shape

Yoav Shoham

Introduction

In the mid-1980s, Yoav Shoham introduced "Naive Kinematics," a formal framework bridging artificial intelligence, geometry, and classical mechanical engineering. Developed as part of the "Commonsense Summer" project at SRI International, this work seeks to formalize how the shapes of physical objects constrain their movement upon contact, providing a logic-based language for common-sense physical reasoning.

Problem Statement: Reasoning About Interaction

Traditional computer representations of shape often treat objects in isolation. However, human intelligence excels at reasoning about interactions—understanding why certain shapes stack, how a piston is constrained by a cylinder, or why a nail resists lateral movement. Shoham addresses the gap in modern kinematics, which often prioritizes complex machinery over the foundational rules of contact. By using first-order predicate calculus, "Naive Kinematics" allows an AI to reason about the "goodness of fit" and "freedom" between interacting shapes.

Methodology: Freedom through Touchpoints

The framework defines the physical interaction of objects in 2D space through the concept of touchpoints.

Translational Freedom

The freedom of an object is defined as the intersection of all motions permitted by its individual touchpoints. When an object contacts an obstacle at a point $P$, that point restricts motion to a specific half-plane or arc. The total translational freedom is the overlap of these allowable directions.

Figure 2: The relationship between a physical configuration and its resulting translational freedom.

Rotational Freedom

Similarly, rotational freedom is defined by identifying points in space that can serve as centers for rotation without causing collisions. The framework uses axioms to determine allowable clockwise or counter-clockwise rotations based on contact constraints.

Key Insights

Freedom vs. Fit

A central contribution of the paper is the distinction between "fit" (how well shapes match) and "freedom" (the range of allowable motion). Shoham illustrates this with a cross inside a square: while the shapes have a "lousy fit," the configuration provides "zero freedom," as the cross cannot move in any direction.

Figure 1: A cross inside a square demonstrates a configuration with poor fit but zero freedom.

The Reuleaux Connection

Shoham's "naive" formulations were discovered to be a generalization of principles established by Franz Reuleaux in the 1870s. While modern engineering shifted toward quantitative calculations, Shoham's work revives these qualitative descriptions for AI applications.

The Local Freedom Paradox

The model identifies a "local" freedom paradox: an object might possess infinitesimal "local" freedom even if it is practically trapped. This suggests the need for a future theory of "continuous freedom" to handle practical physical constraints more effectively.

Key Results

The paper establishes several formal proofs:

  • The Two-Point and Three-Point Theorems: It proves that a minimum of two touchpoints is required to eliminate translation, and three are necessary to eliminate all rotation in a 2D plane.
  • Piston Analysis: The framework correctly models a 2D piston, showing that despite infinite touchpoints along its walls, its freedom is limited to a single axis of translation with zero rotation.
  • Incomparability Theorem: Shoham demonstrates that translational and rotational freedom are independent properties.

Limitations and Future Directions

The current theory is restricted to 2D space, and scaling to 3D presents significant challenges (e.g., requiring at least 7 points to eliminate all freedom). Furthermore, the model is purely kinematic, ignoring dynamics such as friction, mass, and force. Future work is also needed to formalize the "fit" predicate and address the "wiggling" issue inherent in local freedom models.

Summary updated

BibTeX
@inproceedings{shoham1985naive,
  author = {Yoav Shoham},
  title = {{Naive Kinematics: One Aspect of Shape}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1985},
}
Figure