Classify robot failures without inventing a cause

Evidence and interpretation

NIST separates components of robotic performance. Maha’s proposed failure taxonomy uses that separation to locate evidence without making the component name a diagnosis. Start with what the record establishes: pickup did not occur, movement stopped, or the outcome cannot be observed. A missing recording belongs in an evidence-gap category, not automatically in a mechanical-failure category. Keep multiple hypotheses when the available streams cannot discriminate between them. This makes a failure ledger useful to engineers without overstating certainty.

Proposed evidence workflow

1. Record terminal outcome and last observed stage before adding causal labels.
2. Link hypotheses to observations and list the missing test that would discriminate between them.
3. Allow unknown cause and multiple contributing factors; preserve the original label when a later investigation revises it.

Worked illustration

The toy simulation injects grasp-failed as a known software condition. A real empty gripper after pickup would not establish that same cause: the object might never have been reached, might have slipped, or might be outside the camera view. Those are separate investigation branches.

Limits

This is an evidence classification proposal, not a diagnostic controller or a procedure for safely recovering hardware.

Sources and review

Continue reading

• Separate perception evidence from task completion
• Preserve robotics incident evidence without premature conclusions