AI SYSTEM RELIABILITY

Capability Oracles Construction

A core innovation of CanTest is the automated construction of capability-oriented test oracles. These oracles provide independent evaluation metrics for each core capability, enabling precise error detection.

These oracles specifically measure: Perception (Weighted Intersection over Union for object detection), Memory (LLM-evaluated semantic similarity of recalled history), Planning (Normalized Dynamic Time Warping for trajectory alignment), and Decision (Direct comparison of chosen vs. planned actions).

Failure Attribution and Feedback

CanTest's feedback mechanism is designed to pinpoint the exact capability responsible for a task failure, even across long, interdependent task trajectories. It employs counterfactual causal reasoning to identify "failure-inducing errors"—those which, if corrected, would turn a failed trajectory into a success.

This precise attribution, based on identifying the earliest failure-inducing error, generates a comprehensive feedback score for each test case. This score integrates both overall task failure and the severity of capability-specific errors, adaptively guiding subsequent test case generation to focus on the weakest links in the AI agent.

Performance Highlights

Experimental evaluations demonstrated CanTest's significant advantage over state-of-the-art baselines. It consistently discovered a higher number of failure cases and provided more accurate, capability-level diagnoses.

Beyond sheer quantity, CanTest also offers superior qualitative insights. It covers all eight categories of a finer-grained failure taxonomy, including types like "Temporal/Order Error" and "Looping" that baselines missed. The high repair rates (ranging from 81.30% to 96.69%) achieved by correcting identified capability errors validate the fidelity and reliability of CanTest's constructed oracles.

Addressing Limitations

While highly effective, CanTest has areas for future development. A key limitation is its reliance on "expert" models (e.g., optimal routes, semantic annotations) for constructing capability oracles. Future work could explore:

• Leveraging human-in-the-loop supervision for sparse, high-value annotations.
• Training learned surrogate oracles from real logs, distilling "expert" signals from demonstrations and corrective feedback.
• Adapting oracle implementations for direct transfer from simulation to real-world physical environments (sim2real).

These advancements would allow CanTest to provide useful, capability-level failure attribution under increasingly realistic and resource-constrained environments, further enhancing the diagnostic capabilities for embodied AI.