Qualitative Planning with Quantitative Constraints for Online Learning of Robotic Behaviours

Timothy Wiley, Claude Sammut, Ivan BratkoAAAI 2014 · DOI

Abstract

Learning complex robotic behaviors in continuous, noisy state spaces remains a significant challenge. This paper proposes a hybrid control architecture that integrates quantitative constraints into a qualitative planner. By combining the symbolic reasoning of Qualitative Differential Equations (QDEs) with numerical interval arithmetic, the system prunes physically impossible plans and provides a structured framework for behavior refinement. Evaluated on an iRobot Negotiator task, the approach demonstrates improved physical accuracy and the ability to adapt strategies based on environmental scale.

Introduction

Robotic tasks such as navigating urban rubble or climbing stairs often suffer from the "curse of dimensionality" when approached with pure reinforcement learning. Qualitative Reasoning (QR) offers a solution by abstracting continuous variables into symbolic states (e.g., "increasing" or "steady"). However, purely qualitative models lack scale; a planner might treat a 10cm step and a 10m wall as identical obstacles. Historically, addressing this required manual, domain-specific hard-coding. This research introduces Quantitative Constraint Propagation to automate the reasoning of physical limits, such as robot dimensions and obstacle heights.

Methodology

The proposed system utilizes a three-stage Multi-Strategy Architecture to bridge the gap between high-level planning and low-level execution.

The Multi-Strategy Architecture combining qualitative modeling, planning with QSIM, and parameter refinement.
  1. Qualitative Model: The robot and environment are represented symbolically using QDEs.
  2. Qualitative Planner (QSIM): This stage generates parameterized action sequences. The authors' primary contribution is the integration of Interval Arithmetic. By assigning numerical bounds to landmarks (e.g., step height), the planner can mathematically prune search tree branches that represent impossible physical maneuvers.
  3. Numerical Optimization: A trial-and-error learner refines the qualitative plan into exact numerical parameters for real-world execution.

The search process is optimized using A search with heuristics such as MaxQMD (focusing on the variable furthest from the goal) and TotalQMD* (summing qualitative distances across all variables).

Experimental Evaluation

The system was tested using an iRobot Negotiator, a tactical robot equipped with adjustable "flippers" for urban search and rescue. The primary task involved climbing steps of varying heights (10cm to 100cm).

The qualitative model of the iRobot Negotiator, defining variables such as base angle, flipper angle, and positions.

The model tracks several critical variables: robot position ($pos$), flipper angles ($\theta_f$), base angle ($\theta_b$), heading, and velocity ($v$).

Key Results

The integration of quantitative constraints yielded several critical improvements:

  • Physical Realism: The planner correctly identified that simple forward driving suffices for low steps, while higher steps necessitate complex "reverse and lift" maneuvers.
  • Failure Detection: For a 100cm step—exceeding the robot's physical reach—the constrained planner correctly identified that no valid plan existed, whereas an unconstrained planner would have proposed a failing sequence.
  • Heuristic Performance: TotalQMD was found to be more effective for complex goals involving multiple variables, as it captures the overall progress of the system more effectively.

Discussion and Conclusion

While the hybrid approach significantly improves plan validity, it introduces substantial computational overhead, sometimes requiring over 2 billion inferences in Prolog for complex maneuvers. Future work aims to improve efficiency through Answer Set Programming (ASPQSIM) and more sophisticated parameter sampling methods like Markov-Chain Monte Carlo (MCMC). Despite these costs, the ability to use a single qualitative model across diverse environments marks a significant step toward domain-independent robotic learning.

Summary updated

BibTeX
@inproceedings{wiley2014qualitative,
  author = {Timothy Wiley and Claude Sammut and Ivan Bratko},
  title = {{Qualitative Planning with Quantitative Constraints for Online Learning of Robotic Behaviours}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2014},
  doi = {10.1609/aaai.v28i1.9055},
}
Figure