Classifying and Recovering from Sensing Failures in Autonomous Mobile Robots

AAAI 1996

TL;DR

This paper introduces SFX-EH, a novel architecture for autonomous mobile robots to detect, classify, and recover from sensing failures. By leveraging active perception and a "Generate and Test" strategy, SFX-EH can robustly handle unexpected sensor malfunctions or environmental changes, allowing robots to continue their tasks with minimal interruption.

What Problem Does This Paper Address?

Autonomous Mobile Robots (AMRs) are increasingly deployed in complex, real-world environments for tasks ranging from manufacturing to hazardous waste cleanup. A critical bottleneck for their widespread adoption is the need for robust and reliable sensing. If a robot's perception is faulty and goes unnoticed, it can lead to incorrect actions, potentially causing mission failure or even damage. While detecting sensing failures has been explored, the crucial aspect of recovering from these failures, especially when the exact cause is unknown, remains an open research question. This paper tackles this challenge by proposing a systematic approach to not only identify the source of a sensing failure but also to implement a recovery strategy that allows the robot to resume its task.

Methods

The core of the proposed solution is the SFX-EH (Sensor Fusion Effects Exception Handling) architecture. It extends the existing Sensor Fusion Architecture (SFX) by adding a global exception handling module.

The approach has two main phases:

  • Error Classification: This phase employs a novel extension of the Generate and Test strategy, originally developed for expert systems like Dendral. The key innovation here is the integration of active perception. Instead of passively waiting for information, the robot actively uses its sensors to gather additional data to test hypotheses about the sensing failure. This is crucial because AMRs operate in open worlds, making it difficult to pre-model all possible failure modes. The classification process is constrained by the observed symptom (e.g., missing data, low certainty) to reduce the search space.
  • Error Recovery: Once a sensing failure is classified, recovery is designed to be straightforward due to the logical sensor scheme. A logical sensor represents a perceptual process, and the system maintains a knowledge structure (EHKS) that maps failure types to potential alternative logical sensors. Recovery involves searching for and activating a suitable alternative logical sensor. If no alternative can be found, the system gracefully degrades by passing control to the mission planner.

Datasets

Not discussed in the paper. The demonstrations were performed on a Denning MRV-4 mobile robot using its onboard sensors.

Key Insights and Ideas

  • Active Perception is Key for Recovery: The paper strongly emphasizes that for AMRs to recover from unknown sensing failures, they must be able to actively probe their environment and use their sensors to gather more information. This is a significant departure from passive sensing approaches.
  • "Generate and Test" with a Twist: Adapting the classic "Generate and Test" for sensing failures in AMRs, by incorporating active perception and focusing on environmental pre-conditions, makes it more practical and efficient.
  • Logical Sensors Simplify Recovery: Representing perceptual processes as logical sensors, with potential alternatives, creates a structured way to switch to a backup sensing strategy when a failure occurs.
  • Graceful Degradation is Essential: The system acknowledges that not all failures can be recovered from. The ability to gracefully degrade and inform the mission planner is a critical aspect of robust AMR operation.
  • Domain-Specific Knowledge is Necessary but Portable: While SFX-EH relies on domain-specific knowledge about sensor interactions and environmental pre-conditions, the authors argue that this knowledge is more portable than task-specific models, allowing for easier adaptation to new robots or tasks.

Key Results

The paper demonstrates the SFX-EH approach on a mobile robot performing landmark navigation. The system successfully:

  • Detected and Classified Failures: It handled simulated sensor malfunctions (e.g., dirt on lens, disconnected cable), environmental changes (e.g., low light), and errant expectations.
  • Recovered from Failures: In cases of sensor malfunction, it switched to an alternative, redundant logical sensor (e.g., using a rear camera instead of a front one).
  • Resumed Tasks: After recovery, the robot was able to resume and complete its landmark navigation task.
  • Provided Detailed Output: The paper includes an example of the SFX-EH output, showing the step-by-step classification and recovery process.

Limitations and Future Work

The authors acknowledge several limitations and outline future research directions:

  • Domain-Dependent Knowledge: The current SFX-EH relies on domain-specific knowledge for hypotheses and tests. Future work aims to incorporate more general problem-solving strategies and learning mechanisms (e.g., Case-Based Learning) to reduce this dependency.
  • Resource Contention: The current system lacks the ability to resolve resource contention when multiple behaviors demand active perception simultaneously. A global event-driven sensing manager is being developed to address this.
  • Status Updates: SFX-EH does not currently update the sensing status of other behaviors when a failure is detected. This is also a focus for future development.
  • Granularity of Failure Modes: The coarse granularity of failure modes can sometimes lead to slower recovery.
  • Retrying "Bad" Sensors: The paper raises the practical question of when to retry a sensor identified as "bad," which requires further investigation.
  • Bottleneck in Testing: The testing phase, rather than the hypothesis generation, was identified as the bottleneck in recovery. This suggests a need for more efficient testing procedures.
  • Broader Applicability: The authors note that SFX-EH is being applied to intelligent process control for power generation, indicating its potential beyond AMR applications.

Summary updated

BibTeX
@inproceedings{1996classifying,
  title = {{Classifying and Recovering from Sensing Failures in Autonomous Mobile Robots}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1996},
}
Figure