WORLDTECHRePlanMem-VLA Uses Plan Revisions to Decide What a Robot Keeps in Memory
A paper listed on arXiv's Robotics section presents RePlanMem-VLA, a method that lets a vision-language-action policy choose which visual observations to store by watching its own plan revisions. The authors contend that a change of plan can mark the moment task-relevant information has come into view.
How the method decides what to store
According to the arXiv listing, keeping the full history of observations is costly in computation and largely repetitive, which is the case for selective visual memory. RePlanMem-VLA, the abbreviation for Replanning-Guided Keyframe Memory for VLA, makes a policy's plan revisions into a live signal that governs what gets written to memory.
The approach builds on receding-horizon execution, where neighbouring action chunks forecast some of the same future steps. When those time-aligned forecasts diverge by an unusually wide margin, the method reads the gap as a sign that a task-relevant event has occurred.
Reported results in memory-retention tests
The paper reports that the detectors built this way reach ROC-AUC values between 0.981 and 0.997. In the memory-retention experiments it evaluated, RePlanMem-VLA kept task-relevant observations available up to the later decision point in 100% of episodes.
The authors present these findings as evidence that plan revision works as a signal internal to the policy for controlling what a VLA model remembers.
Open questions
- How well does the method perform on tasks and robots beyond the memory-retention experiments evaluated?
- Have the reported results been reproduced independently?