Vision-language-action (VLA) models for robot manipulation typically struggle with events that are only visible over time, such as a brief contact or a temporary occlusion. Looking at just one frame is not enough, but feeding many frames into the model becomes expensive. CondenseVLA, described in a recent arXiv paper, targets this problem by learning to condense the history of observations into a compact set of tokens.

The key idea is to replace the full stack of past frames with a learned, compressed representation that retains the information needed for the current decision. This lets the model make use of multi-frame context without paying the full cost of processing every historical frame at each step. The authors frame this as a form of learnable history condensation, where the model itself decides what to keep and what to discard.

The paper argues that this is especially important for tasks where short-term cues are decisive—like tracking an object being picked up or noticing a change in contact state. By making history efficient, CondenseVLA aims to bring the benefits of temporal reasoning to real-time robot control, where compute budgets are tight. The work is presented as a step toward more practical deployment of multi-frame VLAs in manipulation. No other sources were available for comparison, so the claims here reflect only the paper's own abstract and framing.