Is Your Robot's Memory Just a Clock? Temporal Shortcuts in Memory-Augmented VLA Policies
Is Your Robot's Memory Just a Clock? Temporal Shortcuts in Memory-Augmented VLA Policies
This research paper investigates temporal shortcuts in memory-augmented vision-language-action (VLA) policies. The authors demonstrate that relying on elapsed time rather than event counts leads to 'phantom progress,' where robots incorrectly interpret pauses or null events as task completion. Experiments on the RoboMME benchmark show this vulnerability reduces success rates significantly. The study proposes indexing memory by events instead of clock ticks as an effective solution, which recovers substantial performance without requiring model retraining.
The paper identifies a 'temporal shortcut' in memory-augmented vision-language-action (VLA) policies where reliance on elapsed time rather than event counts causes premature task termination, reducing success rates from 47% to 16% on RoboMME counting tasks; the authors propose indexing memory by events as a fix that recovers over half of the performance loss without retraining.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
robots
1
View original evidence
RoboMME, we make the robot briefly pause or fidget its wrist. Such a null event changes nothing the task depends on, yet a released memory VLA counts the interval as work done and stops early. We call this failure phantom progress. It cuts success from 47 to 16 of 50 on one task and from 44 to 10 on the other. When the same frames are withheld from memory, success returns, so the cause lies in memory rather than in t
Open source S3
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
RoboMME, we make the robot briefly pause or fidget its wrist. Such a null event changes nothing the task depends on, yet a released memory VLA counts the interval as work done and stops early. We call this failure phantom progress. It cuts success from 47 to 16 of 50 on one task and from 44 to 10 on the other. When the same frames are withheld from memory, success returns, so the cause lies in memory rather than in t
Open source S3
arXiv:2611.14488v1 Announce Type: new Abstract: Memory-augmented vision-language-action (VLA) policies are meant to remember what happened. Many of them, however, fill memory by time: they sample frames uniformly across the episode and stamp each frame with its time step. Because time passes even when nothing happens, such a memory may not tell waiting from working. We show that it does not. On the counting tasks of
Open source S2
he interruption itself. We trace the failure to a temporal shortcut: because demonstrations proceed at a steady pace, elapsed time predicts progress. Indeed, rescaling the time stamps alone makes the policy stop early when the clock runs fast and repeat when it runs slow. Surprisingly, removing the time encoding does not cure the failure, because the frames of a null event still occupy slots of the time-sampled memor
Open source S4
y. Across RoboMME's released designs, vulnerability follows how memory is indexed. Of the fixes the diagnosis suggests, training with injected pauses does not remove the failure and impairs counting, whereas indexing memory by events recovers half or more of the loss without retraining. A memory meant to remember what happened should count events rather than clock ticks.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org