Your Robot Doesn’t Need a Bigger Brain. It Needs to Stop Paying Attention to Garbage.
A new study reveals that the real weakness in vision-language-action robots isn’t the model size or the visual encoder — it’s the projector that passes every pixel, noise included. An information bottleneck adapter filters out distractions, boosting robustness by 30% and allowing a tiny 0.5B model to match a 7B one. The future of reliable robots isn’t bigger brains — it’s smarter filtering.