robotics
The Robot Mistook Waiting for Work
A memory-augmented vision-language-action policy counted pauses as progress because its demonstrations made time a shortcut for task state.
Summary
A memory-augmented vision-language-action policy counted pauses as progress because its demonstrations made time a shortcut for task state.
Researchers paused or lightly moved a robot during counting tasks without changing anything the task depended on. The released memory policy nevertheless stopped early: success fell from 47 to 16 of 50 in one task and from 44 to 10 in another, then recovered when the null-event frames were withheld. Changing timestamps alone made the policy stop early or repeat, while removing explicit time encoding did not eliminate the shortcut because time-sampled memory still filled with pause frames. Indexing memory by events recovered at least half the loss without retraining in the reported tests.
Why it matters
A memory-augmented vision-language-action policy counted pauses as progress because its demonstrations made time a shortcut for task state.
Limits and context
- Changing timestamps alone made the policy stop early or repeat, while removing explicit time encoding did not eliminate the shortcut because time-sampled memory still filled with pause frames.
Key claims
A memory-augmented vision-language-action policy counted pauses as progress because its demonstrations made time a shortcut for task state.
Qualification: Changing timestamps alone made the policy stop early or repeat, while removing explicit time encoding did not eliminate the shortcut because time-sampled memory still filled with pause frames.
Evidence: source-2026-11-11-006
Sources
- arXiv:2611.14488v1arXiv / Cornell University · primary research
Corrections
No corrections have been recorded for this story.