Long-horizon, multi-stage vision-language-action (VLA) policies face a problem: the current observation often lacks visual information from earlier stages that is still needed for correct action. The straightforward fix—retaining the entire observation history—is computationally costly, as the paper notes.
The authors propose a replanning-guided visual memory. Rather than storing everything, the model uses replanning signals to decide which past visual information is worth keeping. This targets the moments when the original plan changes and the policy must revisit earlier context.
Because this summary is based on the abstract alone, the specific architecture and experimental results are not described here. The paper is a single arXiv preprint; no independent verification or comparison with other methods is included in the source.