What should a robot tell the person taking over?
When a robot pauses, a person needs more than a status message. A useful AI summary explains what happened and makes the next step easier to understand.
Robotics and AI4 / 4
The robot has stopped. There are images, sensor readings and a long activity log. Someone has to decide what happens next, but the screen first asks them to work out what happened in the first place. This is a familiar problem dressed in new technology: a difficult handover.
I see a useful role for AI here. It can gather the relevant parts and give the reader a place to start. To do that well, it needs to know what the person is trying to decide, rather than simply shorten everything the machine recorded.
Reading guide
Where we are going
A paused inspection gives us a way to ask what the next person really needs to read.
- The reason someone is being asked to help
- The observation behind the explanation
- A summary another person can follow
Tell me why you need me
Imagine an inspection paused because two observations disagree about an aisle. Begin there: the robot needs a review before it can continue. Then show the two observations. A reader should not have to reach the end of the history to find out why they were called.
The next step should also be specific. Is another image needed? Does the responsible operator need to review the route? If it is unclear who can make that choice, the summary should make the gap visible rather than imply the matter is already settled.
Show what the robot saw
“There may be an obstruction” means more when the image and its time are beside it. A measured distance, a camera image and the model’s interpretation can support one another, but the reader still needs to tell them apart.
The NIST AI Risk Management Framework helps connect such questions to the setting in which an AI system is used. In this example, the practical questions are simple: what supports the description, what is missing and who can decide what to do? [1]
A confidence score needs similar care. Guo and colleagues showed why a classifier’s confidence may need calibration before it is interpreted as a probability. In a robotic summary, the number needs an explanation of what it means in that setting. It cannot replace the observation. [3]
Figure 1
The same event, easier to follow
The cards represent observation, interpretation and a human decision. Compare a collection of records, a focused brief and a disagreement that stays visible.
Several records are available, but their presence alone does not tell the reader which decision is waiting or how the observations relate.
The brief gives the reader a clear starting point and keeps the supporting records reachable. It organises the work of review without taking over the decision.
The records do not fully agree. The brief preserves that difference so a person can request another observation or choose an appropriate pause.
A short example of a useful summary
An imagined example: “The inspection is waiting. The rover’s latest image suggests something is blocking the aisle. The overhead image is older and does not resolve the question. A fresh view of the passage is proposed before continuing.”
The paragraph gives a starting point without claiming to settle everything. The images should remain within reach, including the one that disagrees. A useful summary makes it easier to ask a good question, as well as to reach an answer.
Figure 2
Useful anchors for a human-readable brief
| Further reading | The useful idea | The next question |
|---|---|---|
| NIST AI RMF · 2023 | Connects AI risk work to context, measurement and accountable management. | Can the reader identify the risk and who can decide? |
| Guo et al. · 2017 | Shows why a model confidence score needs calibration before probabilistic interpretation. | What does the reported confidence mean in this setting? |
| NODERIQ programme | Frames the relationship between shared evidence, coordination and human authority. | Can a correction be traced into the next handover? |
Keep the correction with the story
The operator may recognise ordinary equipment where the model saw an obstacle. That correction should stay linked to this observation and place. Quietly replacing the original account would make it harder for the next person to understand why the robot stopped.
This is one of the questions behind NODERIQ: how can explanation, review and the original record help the next handover? A useful correction improves the account available to people. Whether it improves a model is a separate matter to test. [2]
Figure 3
A brief that lets the next person get to work
From evidence to a reviewable next step
- State the decision
What needs attention now?
- Show the basis
Keep source, time and disagreement visible.
- Name the next step
Make review and responsibility clear.
Let someone unfamiliar read it
My simplest test is to ask someone who was not there to read the summary. Can they explain why work paused, find the supporting observation and say what still needs a decision? If they first need a colleague to translate the language, it needs another edit.
Two versions made from the same records could be compared this way. Keep the information constant and look at where readers misunderstand. That would teach us more than asking which text sounds more professional. The point of the brief is to make the next person’s work easier.
Sources
- NIST: AI Risk Management Framework nist.gov
- Neura Parse: NODERIQ Research Programme neuraparse.com
- Guo et al.: On Calibration of Modern Neural Networks proceedings.mlr.press