Memory VLAs treat pauses as progress via temporal shortcuts
Original title: Is Your Robot's Memory Just a Clock? Temporal Shortcuts in Memory-Augmented VLA Policies
Brief pauses drop counting success from 47 to 16 of 50 and from 44 to 10; event-indexed memory recovers half or more without retraining.
The paper shows memory-augmented vision-language-action policies that sample frames uniformly and stamp them with time treat null events as work. On RoboMME counting tasks, a brief pause or wrist fidget makes a released memory VLA stop early, a failure called phantom progress. Withholding those frames restores success, and rescaling time stamps alone changes when the policy stops or repeats. Training with injected pauses does not remove the failure, while indexing memory by events recovers half or more of the loss without retraining.
- Success falls from 47 to 16 of 50 on one counting task.
- The other task falls from 44 to 10 of 50.
- Removing time encoding does not cure the failure.
- Vulnerability tracks how memory is indexed across released RoboMME designs.
From the original article
1 min read at huggingface.co
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.
By Chubin Zhang, Bo An · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.