Researchers have found a way to make robot control software smarter about how far ahead it plans, without retraining it.
Most vision-language-action systems that guide robots use "action chunking": instead of predicting one move at a time, they predict a batch of future actions at once, which is faster but less responsive to surprises. Today that batch size is usually fixed, whether the robot is doing something easy like moving through open space or something finicky like gripping a small object. A new method called GeoAAC, described in a paper posted to arXiv, changes the batch size on the fly by examining the internal math of the prediction process itself. It reads geometric patterns in the model's denoising steps, borrowed from flow matching, to gauge how confident a given prediction is, then lengthens or shortens the action horizon accordingly. Tested on two existing robot policies, GR00T N1.5 and pi0.5, across several simulation benchmarks and real-world manipulation tasks, it beat fixed-horizon baselines and other adaptive methods, including a jump in real-world task success from 53.3% to 74.4%.
That real-world number is the one that matters. Simulation gains in the single digits are common in robotics papers and often evaporate outside the lab; a 21-point jump on physical manipulation tasks is a different order of claim, and it comes from a technique that needs no additional training data or fine-tuning, just smarter use of information the model already produces.
If the result holds up under independent testing, it is a reminder that robotics progress right now is less about bigger models and more about squeezing more judgment out of the ones that already exist.