Generating Perception Requests and Expectations to Verify the Execution of Plans

Richard J. Doyle, David Atkinson, Rajkumar DoshiAAAI 1986

In the mid-1980s, as robotics moved toward complex, real-world applications like satellite repair, researchers at NASA’s Jet Propulsion Laboratory (JPL) identified a critical flaw in classical AI planning. While planners could generate sequences of actions, they often operated under a "closed-world assumption," presuming that every command would be executed perfectly. In reality, sensors are noisy and actuators can fail. To bridge this gap, the authors presented GRIPE (Generator of Requests Involving Perceptions, and Expectations) at AAAI-86, a system designed to make robots "aware" of their own execution success.

The Challenge of Execution Monitoring

Classical planners typically assume that if a plan dictates an action, the world state updates accordingly. However, for a robot performing a 67-step repair on the Solar Max satellite, a single missed grasp or a slipped gear could lead to mission failure. Execution monitoring—the process of verifying that a plan is proceeding as intended—is essential.

Manually coding sensor checks for every step of a complex plan is impractical. The authors focused on the Generation problem: automatically determining which sensors to use and what values to expect to verify that a plan's preconditions and postconditions are met.

The GRIPE System

GRIPE functions as a pre-processor that augments symbolic robot plans with specific perception requests. At its core are Verification Operators, which map symbolic assertions to physical sensor data.

Structure of a verification operator in GRIPE.

These operators allow the system to link a high-level goal, such as "the object is at location X," to a concrete perception request, such as "use the vision system to SEE the object." GRIPE also employs Intent-Based Verification, recognizing that the reason for an action determines the required precision. A delicate grasp requires multi-modal verification (vision and force), while a simple arm retraction might only need a cursory position check.

Verification in Practice

The system was tested using a scenario based on the Solar Max satellite repair. GRIPE successfully transformed "blind" plans into "aware" ones by integrating four primary sensor types: arm-kinesthetic (position), hand-kinesthetic (gripper width), force sensors, and vision systems.

Mapping of sensors to specific perception requests.

By analyzing the preconditions and postconditions of each action using situation calculus, GRIPE inserts perception requests at critical junctures. It also handles Recursive Subgoaling; if a sensor requires a specific condition to function—such as a camera needing a clear line of sight—GRIPE can generate a subgoal for the planner to reposition the robot.

Key Insights: Perception vs. Inference

A central tenet of the paper is that perception-based checks are inherently more reliable than symbolic inferences. The authors introduce a formal distinction between MODEL= (comparing internal beliefs) and SENSE= (comparing real-time perception against the internal model). This distinction is the "essence of verification."

Furthermore, the authors treat sensing as a finite resource. Because perception takes time and computation, it must be planned and scheduled. GRIPE uses dependency pruning to avoid redundant checks—if a condition is verified as a postcondition of one action and immediately serves as a precondition for the next, it is only checked once.

Conclusion and Future Directions

While GRIPE represented a significant step forward, the authors noted several areas for improvement. The prototype primarily focused on "snapshots" of verification rather than continuous monitoring, which is necessary for tasks like maintaining a grip on a heavy object during transit. Future work suggested the need for "Importance Criteria" to prioritize monitoring based on the criticality of specific plan assertions, as well as tighter integration between the planner and the monitoring system to allow for dynamic communication during execution.

This foundational work highlighted that for autonomous systems to be truly reliable, they must not only plan their actions but also plan how they will perceive the success of those actions.

Summary updated

BibTeX
@inproceedings{doyle1986generating,
  author = {Richard J. Doyle and David Atkinson and Rajkumar Doshi},
  title = {{Generating Perception Requests and Expectations to Verify the Execution of Plans}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure