Vision-Language-Action (VLA) models offer a promising path to general-purpose robot manipulation, but their closed-loop deployment is often hamstrung by inference efficiency. At every timestep, the model must process high-resolution, often multi-view images, which creates a computational bottleneck that slows down real-time control.

FastVLA, described in a new arXiv preprint, tackles this by introducing task-conditioned and temporally adaptive visual token compression. Instead of feeding all visual tokens to the model at every step, the method selectively compresses them based on the current task and the temporal context, reducing the number of tokens that need to be processed.

The paper's abstract is truncated, so detailed experimental results are not available in the source. Still, the core idea is clear: making VLA models faster by being smarter about which visual information matters at each moment, rather than brute-forcing through every pixel. This is a practical step toward real-time robot control with large multimodal models.