Abstract
This paper presents a hybrid approach to robotic control that integrates high-level Qualitative Reasoning (QR) with low-level trial-and-error learning. By leveraging Qualitative Simulation (QSim) within a Multi-Strategy Learning (MSL) framework, the proposed system enables robots to generate general action plans based on "common sense" physics. This allows for adaptation to complex, noisy environments without the need for task-specific planners. Preliminary results using the iRobot Negotiator platform demonstrate successful qualitative state-transition mapping and efficient path generation for traversing rough terrain.
Introduction
Robotic planning in real-world environments is challenged by continuous state spaces, noise, and unpredictability. Traditional approaches often fall into two categories: specialized model-based reinforcement learning, which lacks flexibility, and behavioral cloning, which struggles to generalize. This research addresses the "specialization gap" by proposing that robots should reason about environmental physics qualitatively—similar to how humans understand that certain actions produce specific physical trends—before refining these behaviors through practice.
Methodology
The core of this approach is the integration of Qualitative Simulation (QSim) into a Multi-Strategy Learning (MSL) architecture. MSL mimics human skill acquisition, starting with high-level instructions (the plan) and progressing to fine-tuned execution (the "feel").
The process involves several key stages:
- Qualitative Modeling: Variables are described by their relationships to landmarks (e.g., "increasing" or "at zero") rather than precise coordinates. This abstraction allows the planner to ignore the infinite complexity of continuous space.
- Qualitative Planning: QSim generates a directed graph of possible qualitative state transitions. The system identifies a path through this graph from the current state to the goal.
- Parameter Refinement: Once a high-level plan is generated, a constraint solver bounds the parameters, and a trial-and-error learner identifies optimal quantitative values.
Experimental Setup
The framework was evaluated using the iRobot Negotiator, a tracked robotic platform equipped with adjustable "flippers." The primary task involved traversing rough terrain, such as climbing steps and obstacles, where the relationship between flipper movement and base pitch is critical for stability and progress.
Results and Discussion
Preliminary modeling demonstrated that the Negotiator's movements can be accurately mapped into a qualitative state-transition graph. The system successfully identified sequences of qualitative states to achieve a desired pitch, which were then converted into parameterized actions. Referencing prior work, this MSL approach suggests that complex behaviors can be optimized in significantly fewer trials (e.g., ~40) compared to pure reinforcement learning.
However, several challenges remain. QSim's non-deterministic nature can lead to multiple predicted outcomes for a single action, complicating path selection. Additionally, the boundary between qualitative and quantitative reasoning needs refinement to distinguish between obstacles of different scales (e.g., a small step vs. a wall). Future work will focus on more efficient search methods to handle the exponential growth of state-transition graphs as more variables are introduced.
Conclusion
By separating high-level qualitative planning from low-level quantitative refinement, this research provides a promising path toward more general and adaptable robotic systems. The use of QSim as a "physics engine" allows the framework to be applied across diverse robotic configurations, provided a qualitative model exists.
Summary updated