Sensible Planning: Focusing Perceptual Attention

Lonnie Chrisman, Reid G. SimmonsAAAI 1991

Abstract

In real-world robotics, sensing is not a free resource; it consumes time, power, and computational cycles. This paper addresses the challenge of rational perception by introducing "static sensing policies." By committing to a fixed sensing strategy for a specific task, the authors simplify the computationally expensive Partially Observed Markov Decision Process (POMDP) into a standard Markov Decision Process (MDP). This framework allows robots to balance the cost of information against its utility, leading to more efficient and robust planning.

Introduction

Classical AI planning often assumes perfect environmental knowledge, while reactive robotics assumes continuous, omniscient monitoring. Neither extreme is practical for complex tasks where sensors are limited and expensive. When a robot must choose between thousands of visual routines, it faces a "state-space explosion" if it treats every sensing decision as a new planning state. This paper proposes a middle ground: treating sensing as a strategic choice rather than a constant background process.

Methodology: Static Sensing Policies

The core innovation is the Static Sensing Policy, which structures how a robot interacts with its environment through a "sense-act" cycle.

1. Sensing Procedures

Instead of deciding what to sense at every millisecond, the robot employs a "sensing procedure"—a decision tree of sensor checks that results in a specific distinction about the world state.

2. Folding Costs into MDPs

By fixing the sensing policy for the duration of a task, the costs associated with perception can be integrated directly into the reward structure of an MDP. This transformation enables the use of standard, efficient algorithms like Howard’s policy iteration to find optimal plans without the complexity of belief-state reasoning.

The structure of static sensing policies (a) and randomized choices (b) within the sense-act cycle.

3. Mega-Actions and Hierarchy

To overcome the rigidity of a static policy, the authors introduce "mega-actions." These are hierarchical sub-plans that can possess their own internal sensing policies. This allows a robot to adapt its focus as it progresses through different stages of a complex task, effectively changing what it pays attention to as needed.

Experimental Evaluation: The Robot and Cup Domain

The proposed framework was evaluated using the "Robot and Cup Domain," a toy problem designed to test the trade-offs between sensing cost and task success.

The experimental setup for the Robot and Cup Domain.

In this domain, the goal is to grasp a cup that can be in several states: Upright ($S_U$), Tipped Forward ($S_F$), Tipped Backward ($S_B$), or Grasped ($S_G$). The robot has two sensors: $SO_1$, which distinguishes if the cup is upright or tipped, and $SO_2$, which identifies the specific direction of the tip. Crucially, $SO_2$ is significantly more expensive than $SO_1$, forcing the robot to decide if detailed orientation data is worth the cost.

Key Results and Insights

The study compared several strategies, measuring the "gain" (average reward per step):

  • Full Sensing: Utilizing both sensors provided maximum information but high overhead, resulting in a gain of -0.96.
  • Limited Sensing (Randomized): Allowing the robot to make randomized choices when states were indistinguishable (due to limited sensing) improved the gain to -0.72.
  • Mega-Actions: The hierarchical approach, which combined specific actions like spinning and grasping into a robust procedure, outperformed all others with a gain of 0.619.

These results highlight that more sensing is not always better. A well-designed procedural plan can often compensate for limited perception, avoiding the high costs of constant, detailed environmental monitoring.

Conclusion

This research demonstrates that perception should be treated as a rational, planned activity. By focusing attention through static sensing policies and hierarchical mega-actions, robots can achieve complex goals more efficiently. Future work will focus on scaling these methods to larger state spaces and exploring how robots might autonomously learn or cache these sensing procedures from experience.

Summary updated

BibTeX
@inproceedings{chrisman1991sensible,
  author = {Lonnie Chrisman and Reid G. Simmons},
  title = {{Sensible Planning: Focusing Perceptual Attention}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1991},
}
Figure