Planning Robot Control Parameter Values with Qualitative Reasoning

Stephen F. Peters, Shigeoki Hirai, Toru Omata, Tomomasa SatoIJCAI 1991

Abstract

This study explores a hybrid approach to robotic control that bridges symbolic AI and physical execution. By combining qualitative reasoning (QR) with numerical history, the proposed system addresses the brittleness of traditional numerical models in uncertain environments. The system was tested on a delicate match-striking task, demonstrating that qualitative physics can provide a robust heuristic for narrowing down control parameters such as force, velocity, and angle, leading to successful task completion in a minimal number of trials.

Introduction

In robotics, numerical models are often "brittle." While a robot might possess a perfect geometric map, minor errors in calibration, friction, or material properties can lead to task failure. Traditional solutions often involve expensive sensors or complex error-recovery code. This paper, published in the early 1990s, proposes an alternative: a hybrid planner that uses qualitative physics to guide a robot's search for the correct control parameters, effectively providing "common sense" to the numerical search process.

Methodology and Experimental Setup

The researchers integrated their planner into the MEISTER (Model Enhanced Intelligent and Skillful TEleoperational Robot) system. The core task involved using a 6-DOF ETA2 slave manipulator to strike a match—a process requiring precise control over force, velocity, and angle to ensure ignition without structural failure.

Physical parameters for the match-striking task including force, velocity, and angle.

Qualitative Reasoning and Likelihood Potentials

The system utilizes a Qualitative Reasoning Network where nodes represent physical properties (e.g., bending moment, heat accumulation) and edges represent qualitative relationships. A significant challenge in qualitative math is ambiguity—for instance, determining the result of a "positive" value plus a "negative" value. To resolve this, the authors introduced "likelihood potentials," which assign relative probabilities to outcomes, allowing the planner to select the most promising direction for parameter adjustment.

Qualitative addition table illustrating the handling of ambiguity in qualitative reasoning through likelihood potentials.

Learning from Execution History

The system maintains a history of every trial. If a match breaks at a specific "digging depth" (a proxy for force), that value is marked as a numerical limit. These limits are propagated back through the qualitative network to prune the search space for subsequent trials. Instead of a blind numerical search, the QR network generates a "qualitative vector"—a suggested direction in the parameter space, such as "decrease angle and increase depth."

Key Results

The planner demonstrated high computational efficiency, requiring less than 3 seconds per trial on a Sparc station. In a documented match-striking experiment, the robot failed its first 10 trials due to non-ignition or breakage. However, by the 11th trial, the system had successfully narrowed the search space to find a reliable combination: an angle of -50 degrees and a digging depth of 10mm. The qualitative guidance effectively transformed a wide range of numerical possibilities into a small, reliable goal region.

Conclusion and Future Work

The integration of qualitative reasoning provides a powerful heuristic for robotic parameter tuning under uncertainty. While the current system requires manual modeling of the physical relationships, it proves that "fuzzy" physical logic can drive precise motor control. Future research directions include automating the discovery of physical models and applying this hybrid approach to more complex, multi-dimensional manipulation tasks.

Summary updated

BibTeX
@inproceedings{peters1991planning,
  author = {Stephen F. Peters and Shigeoki Hirai and Toru Omata and Tomomasa Sato},
  title = {{Planning Robot Control Parameter Values with Qualitative Reasoning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1991},
}
Figure