A robot can follow instructions perfectly…and still fail.
“Pick up the bottle.” It does. The ketchup spills.
The robot did what you asked, but not what you needed. The instruction missed what mattered.
Hallucination is different: the system invents an observation or claims a task is complete when it isn’t.
One needs clearer goals. The other needs evidence. Both need the actual outcome checked.
That’s the idea behind Proof of Perception: check claims against independent, live sensor observations, not just one camera, model, or robot’s word.
Check what was there before, what happened during, and what changed after. When evidence is missing or conflicting, don’t declare success.
Following instructions isn’t the same as finishing the job.
And saying “done” isn’t proof.
"If you can’t be right, at least be sure when you’re wrong"
Errors are always going to happen in robotics but they need to meet 3 criteria to be acceptable -
1. Safe to the robot & people
2. Minimal intervention to correct
3. Minimal false positives/negatives
Deep learning often fails at number 3.
LLMs frequently hallucinate incorrect results & find a post hoc justification for why the results are correct.
This isn't ideal in the world of words and it's disastrous in the physical world of robots.
Is there a route to creating AI robots that don't hallucinate?