Using Platform Models for a Guided Explanatory Diagnosis Generation for Mobile Robots

Daniel Habering, Till Hofmann, Gerhard LakemeyerIJCAI 2021 · DOI

TL;DR

This paper introduces a novel approach to help mobile robots understand and recover from failures during task execution. By modeling the robot's hardware components as state machines, the system can generate more accurate explanations for failures and suggest appropriate recovery actions, significantly improving robot performance in complex scenarios.

What Problem Does This Paper Address?

Mobile robots operating in the real world are prone to errors. These errors can stem from unexpected environmental changes or, more critically, from internal hardware malfunctions. When a failure occurs, the robot's internal understanding of the world state can become inconsistent, making it difficult to continue its task or even to understand why the failure happened. Traditional diagnosis methods often treat action failures as random events, lacking the ability to pinpoint the underlying cause. This paper addresses the challenge of generating explanatory diagnoses for robot failures, specifically focusing on hardware-related issues, and using these diagnoses to guide recovery and enable continued operation.

Methods

The core of the proposed approach lies in integrating a detailed model of the robot's hardware platform into the diagnosis process. Here's a breakdown of their methods:

  • Hardware Platform Modeling: Robot components (like navigation or localization units) are modeled as finite state machines. These models capture normal operation, potential fault states (e.g., "stuck," "broken"), and the transitions between them, often triggered by exogenous events (environmental influences). Probabilities are assigned to exogenous transitions to aid in diagnosis.
  • Diagnosis as Planning: The diagnosis problem is framed as a planning problem using PDDL (Planning Domain Definition Language). This allows leveraging existing, powerful planning algorithms.
  • Action Variants and History Enforcement: To explain failures, the system generates "action variants" that represent how original actions behave under faulty component states (e.g., a "move-stuck" action). Crucially, the PDDL domain is constructed to enforce the robot's action history, meaning any generated explanation must be a plausible adaptation of the original sequence of actions, not just any sequence that explains the observation.
  • Top-k Planning: A top-k planner is used to generate a set of the k most cost-optimal diagnoses. The cost metric prioritizes explanations with fewer exogenous actions, aligning with Ockham's razor principle.
  • Active Diagnosis: To resolve ambiguity among multiple diagnosis candidates, the system employs active diagnosis. It strategically selects sensing actions that are most likely to differentiate between remaining explanations, using the probabilities from the hardware models to maximize information gain.
  • Recovery: Once a unique explanation is found, the robot updates its world model and attempts to repair the identified faulty component. If repair is not possible, it may disable itself or request human assistance.

Datasets

Not discussed in the paper. The evaluation uses a simulated RoboCup Logistics League (RCLL) scenario.

Key Insights and Ideas

  • Hardware is Key to Explanations: The most significant insight is that understanding why a robot fails requires modeling its internal hardware states. Treating failures as random events limits recovery options; knowing a gripper is "decalibrated" is far more actionable than just knowing a "put" action failed.
  • Diagnosis as Planning Unifies Concepts: Framing diagnosis as a planning problem elegantly integrates action history, action variants, and exogenous events into a single framework, leveraging mature planning technology.
  • History Matters for Plausibility: Forcing explanations to adhere to the original action history is crucial for generating realistic and actionable diagnoses, preventing the planner from inventing entirely new, unrelated sequences.
  • Active Diagnosis with Probabilistic Models: Using probabilities from hardware state transitions to guide active diagnosis allows for more informed and efficient selection of sensing actions, directly linking engineering knowledge to diagnostic intelligence.

Key Results

The paper evaluates its approach in a simulated RoboCup Logistics League (RCLL) scenario. Key findings include:

  • Increased Success Rate: Robots using diagnosis (especially with platform models) had a higher percentage of successful plans compared to those without diagnosis.
  • Platform Models Outperform: Diagnosis with platform models significantly outperformed diagnosis without platform models. This is attributed to the ability to identify and repair underlying causes, rather than just observing random failures.
  • Improved Game Scores: In configurations with frequent failures, diagnosis with platform models led to significantly higher team game scores.
  • Comparable Runtime: The computational cost (runtime) of diagnosis with platform models was comparable to diagnosis without platform models, suggesting the added modeling complexity doesn't incur a prohibitive performance penalty.

Limitations and Future Work

  • Assumption of Independent Faults: The current approach assumes faults are independent. Future work could explore how to handle correlated or cascading failures.
  • Perfect Sensing Assumption: For active diagnosis, the paper assumes perfect sensing. Extending this to noisy sensors would be a valuable next step.
  • Infinite Explanations: The paper acknowledges that there can be an infinite number of possible explanations and relies on the top-k planner to limit the search space. Further research could explore more sophisticated methods for managing this complexity.
  • Generalizability: While evaluated in an RCLL scenario, generalizing the approach to a wider range of robot platforms and tasks would be beneficial.

Summary updated

BibTeX
@inproceedings{habering2021using,
  author = {Daniel Habering and Till Hofmann and Gerhard Lakemeyer},
  title = {{Using Platform Models for a Guided Explanatory Diagnosis Generation for Mobile Robots}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/263},
}
Figure