Semantic-Based Explainable AI: Leveraging Semantic Scene Graphs and Pairwise Ranking to Explain Robot Failures

Devleena Das, Sonia Chernova

Introduction

As robots transition from controlled industrial settings to unstructured domestic environments, failures in tasks like object manipulation are inevitable. For these robots to be useful, they must be able to explain their failures to "everyday users"—non-experts who can then assist in recovery. Traditional explanation methods often rely on hand-annotated context that fails to scale or provide "context-based" (CB) explanations that are too vague (e.g., "the object is occluded") to be actionable. This paper proposes a framework that generates semantically descriptive explanations by leveraging Semantic Scene Graphs (SSGs) and a pairwise ranking algorithm to identify the most relevant failure causes.

Methodology: Semantic Scene Graphs and Relevance Ranking

The proposed framework operates in three stages to transform raw visual data into human-readable explanations:

  1. Semantic Scene Graph (SSG) Generation: Using an adapted MOTIFNET model, the system generates a graph where nodes represent objects and their attributes (e.g., "fragile," "heavy"), and edges represent spatial relationships (e.g., "underneath," "close to").
  2. Pairwise Ranking (Relevance Filter): Raw scene graphs are often cluttered with irrelevant information. To filter this, the authors employ a Pairwise Ranking Algorithm using Random Forest classifiers. This step determines which specific relationship or attribute is the primary cause of failure, effectively placing the robot's "attention" on the most pertinent data.
  3. Natural Language Generation: The highest-ranked semantic triples are inserted into templates to produce explanations such as, "I could not pick up the cup because it is behind a large bottle."

The researchers categorized failures into a specific taxonomy to test the system's robustness across different scenarios, ranging from simple spatial constraints to complex attribute-based refusals.

Taxonomy of failure causes (Spatial vs. Attribute) and their mapping to failure types (Single, Compound, Attribute).

Experimental Evaluation

The framework was evaluated using 188 cluttered household scenes from the AI2Thor simulator. A user study involving 90 participants compared three conditions: no explanation ("None"), the previous state-of-the-art "Context-Based" (CB) explanations, and the proposed "Semantic-Based" (SSG-R) explanations. Participants evaluated four failure categories: Single Spatial, Compound Spatial, Attribute, and Aggregate failures.

Key Results and Discussion

The results demonstrate that semantic-based explanations significantly outperform both the baseline and context-based methods in terms of perceived helpfulness and the user's ability to identify the failure cause and solution.

Perceived helpfulness across (a) Aggregate, (b) Single, (c) Compound, and (d) Attribute failures, showing the significant advantage of Semantic-Based explanations.

Key findings include:

  • Statistical Significance: For compound and attribute failures, the SSG-R method outperformed all other conditions with high statistical significance ($p < 0.001$).
  • The Importance of Ranking: Without the pairwise ranking step, explanations often included extraneous details that confused users. Ranking ensures the explanation is concise and relevant.
  • Actionable Intelligence: By moving beyond vague labels like "occluded" to specific semantic descriptions (e.g., "occluded by a glass bottle"), the robot provides the user with the necessary information to intervene effectively.

Conclusion and Future Work

This work highlights that for robot explanations to be effective, they must be both semantically grounded and filtered for relevance. While the current model achieved 84.9% accuracy in relationship classification, future work aims to expand the vocabulary to "open-world" scenarios and apply the framework to other failure modes, such as navigation errors.

Summary updated

BibTeX
@inproceedings{das2021semantic,
  author = {Devleena Das and Sonia Chernova},
  title = {{Semantic-Based Explainable AI: Leveraging Semantic Scene Graphs and Pairwise Ranking to Explain Robot Failures}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2021},
  doi = {10.1109/iros51168.2021.9635890},
}
Figure