Spatial and Qualitative Aspects of Reasoning about Motion

Kenneth D. ForbusAAAI 1980

Abstract

Kenneth Forbus introduces FROB, a system designed to reason about motion in a "Bouncing Ball" world by bridging the gap between precise geometry and qualitative common sense. By combining numerical "Metric Diagrams" with symbolic "Space Graphs," the paper demonstrates how AI can predict trajectories, detect physical inconsistencies, and answer complex questions like "Can these two balls ever collide?"

Introduction: The Challenge of Qualitative Physics

In 1980, AI research faced a dichotomy: systems either employed rigid formal mechanics involving complex differential equations or relied on purely symbolic logic that struggled with the nuances of physical space. Humans, however, reason about motion fluently without performing calculus. We possess an intuitive understanding that a weakly thrown ball will not escape a deep well.

A sample trajectory of a point-mass ball within the FROB environment.

Forbus addresses this through Qualitative Spatial Reasoning. The goal is to build a program that understands the "common sense" of motion—recognizing types of movement (flying vs. colliding) and the topology of a scene (inside vs. outside)—while maintaining enough mathematical grounding to ensure physical validity.

Methodology: The FROB Architecture

FROB operates in a 2D world of point-mass balls and line-segment surfaces. Its methodology relies on three core representations:

1. The Metric Diagram

This hybrid representation stores both symbolic elements and numerical coordinates. It acts as a "digital diagram" that the program queries to determine spatial relationships, such as which side of a line a point resides on.

2. The Space Graph

FROB discretizes the empty space in a diagram into a vocabulary of PLACES (regions, lines, points). This graph defines connectivity (e.g., Region A is ABOVE Region B), providing a framework for qualitative reasoning.

Mapping physical space (Geometry of Space Graph) to a symbolic network (Pointer Structure).

3. Action Sequences & Envisioning

Motion is described through Action Sequences, which are networks of "acts" (FLY, COLLIDE, SLIDE) and "states." Through a process called Envisioning, FROB generates a Sequence Graph—a map of every possible qualitative state a ball could reach from its current position.

Key Insights and Ideas

  • The Visual Apparatus Conjecture: Forbus suggests that human spatial fluency stems from our visual system's ability to simplify space into discrete, meaningful regions rather than through theorem proving.
  • Finite Qualitative States: Despite infinite numerical coordinates, there are only a small number of qualitative states (e.g., "Flying Up in Region 1"). This makes envisioning every possible future computationally feasible.
  • Pruning via Global Assumptions: FROB can assimilate high-level constraints—such as perfect elasticity or energy conservation—to prune the Sequence Graph, ruling out impossible motions without full step-by-step simulation.

Results and Performance

FROB successfully answers four fundamental types of spatial queries:

  1. Immediate Next Action: Will the ball hit a wall or continue flying?
  2. Next Location: Which region will the ball enter next?
  3. Terminal Behavior: Will the ball eventually stop or become trapped?
  4. Collision Detection: Do the Sequence Graphs of two balls overlap in the same PLACE at the same time?

The system effectively detects physical inconsistencies, such as a ball bouncing higher than its starting point, and can summarize long-term behavior in complex environments like "wells."

Limitations and Future Work

While pioneering, FROB relies heavily on quantitative parameters for inconsistency detection—a task humans often perform with less information. The 1980 model also ignores ball shape, spin, and air resistance. Future research directions include expanding the system to 3D space, handling multi-object interactions, and moving toward even more purely qualitative methods for parsing motion data.

Summary updated

BibTeX
@inproceedings{forbus1980spatial,
  author = {Kenneth D. Forbus},
  title = {{Spatial and Qualitative Aspects of Reasoning about Motion}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1980},
}
Figure