Rational-Based Visual Planning Monitors

Zohreh AlaviIJCAI 2016

Introduction

In traditional robotics, planning and perception are often treated as distinct, sequential modules. A planner typically assumes a static world state while calculating a sequence of actions, while the vision system provides periodic updates without understanding the context of the robot's goals. This separation becomes problematic in dynamic environments where conditions can change rapidly. If an environmental change renders a current plan obsolete—such as a fire being extinguished while a robot is still planning how to fetch an extinguisher—a standard system may waste significant computational resources completing a plan that is no longer necessary.

This paper introduces a method to bridge this gap by linking visual perception directly to the "rationales" behind planning decisions. By implementing rationale-based monitors within a Hierarchical Task Network (HTN) planner, the system can detect when environmental shifts invalidate current planning logic, allowing for immediate pivot and adaptation.

System Architecture and Methodology

The proposed framework integrates two primary cognitive components: the Simple Hierarchical Ordered Planner (SHOP) and the Metacognitive Integrated Dual-Cycle Architecture (MIDCA).

1. HTN Planning with SHOP

Unlike classical planners that search for a linear sequence of moves, SHOP utilizes Hierarchical Task Networks to decompose high-level goals (e.g., "clear a table") into smaller subtasks. This process continues until the tasks are reduced to primitive actions (e.g., "pick up block A").

2. Rationale-Based Monitors

The core innovation lies in the creation of monitors based on planning rationales. When the planner selects a specific action or decomposition based on a condition in the environment (e.g., "I am fetching the extinguisher because the block is on fire"), the system generates a monitor for that specific condition.

3. Visual Integration and Plan Transformation

These monitors serve as specific instructions for the vision system, directing it to watch for changes in "load-bearing" environmental features. If a monitor detects that a condition has changed (the fire is out), it "fires," triggering a plan transformation. Rather than restarting from scratch, the system backtracks to the specific decision point influenced by that rationale and refines the plan based on the new reality. This allows the robot to skip unnecessary steps, such as fetching an extinguisher for a fire that no longer exists.

Experimental Evaluation

The researchers evaluated the system using a Modified Blocksworld Domain implemented on a Baxter humanoid robot using the Robot Operating System (ROS).

In this simulated environment, blocks can catch fire. To safely interact with a burning block, the robot must locate and use an extinguisher, which may be hidden within a tower of other blocks. The experiments varied two primary factors:

  • Task Complexity: Tower heights ranged from 4 to 26 blocks to test the scalability of the planning logic.
  • Temporal Dynamics: The timing of environmental changes (e.g., the fire going out) was varied across 0, 10, 30, or 50 planning steps.

Key Insights and Results

Efficiency and Selective Attention

Vision is computationally expensive. By using planning rationales to guide perception, the robot practices a form of "selective attention," only spending resources to monitor features that are critical to its current strategy. This targeted approach prevents the system from being overwhelmed by irrelevant environmental data.

The Value of Early Interruption

The experiments demonstrated that rationale-based monitors significantly reduced total planning time when the environment changed. The most substantial savings occurred when changes happened early in the planning process, preventing the "sunk cost" of completing a complex, useless plan. However, even when changes occurred late, the monitors provided value by preventing the system from finalizing and attempting to execute an invalid plan.

Scalability

As the complexity of the task increased (i.e., taller towers), the benefits of rationale-based monitoring became more pronounced. In high-complexity scenarios, the cost of "wrong" planning is significantly higher, making the ability to interrupt and adapt mid-process essential for efficient operation.

Conclusion and Future Work

This research proves that rationale-based monitors, originally conceived for state-space planners, are highly effective when applied to the hierarchical decomposition of HTN planning. By allowing the planner to interrupt itself during the reasoning phase, the system achieves a higher level of fluid intelligence in dynamic settings.

Future work will focus on extending these monitors into the execution phase, enabling the robot to respond to unexpected changes while physically performing actions. Additionally, the authors aim to implement parallel processing for the visual interpreter to ensure that the robot's perception and reasoning cycles are fully synchronized, further reducing latency in high-stakes environments.

Summary updated

BibTeX
@inproceedings{alavi2016rational,
  author = {Zohreh Alavi},
  title = {{Rational-Based Visual Planning Monitors}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
}
Figure