PaceVLA research reports faster robot execution while preserving task success

Researchers report that PaceVLA improved robot execution efficiency while maintaining task success rates in three multi-task simulation benchmarks and six real-world tasks. The arXiv preprint describes tests across multiple robot embodiments, including long-horizon and dexterous manipulation. PaceVLA selects between actions with different levels of timing detail at each inference step. It combines action decoders supporting different execution intervals and prediction horizons with a mechanism that judges when coarser actions are appropriate. That selection addresses a tradeoff: combining low-level actions can accelerate movement but discard details needed for precise interactions.