A robot can receive a command based on where an object was, not where it is now: the object keeps moving while the AI works out what to do. An October 1 revision of a January preprint from Nanyang Technological University researchers describes DynamicVLA, a robot-control model designed to account for that delay.
DynamicVLA is a vision-language-action (VLA) model: it combines camera images and spoken or written instructions to produce robot movements. It starts its next round of reasoning as soon as the previous one finishes, while the arm continues moving. When a new set of actions arrives, the system drops the commands whose intended execution times have already passed and uses the remaining steps. That keeps a plan tied to the moving scene instead of replaying instructions for an earlier moment.
In the authors’ simulated benchmark, DynamicVLA completed 47.06% of trials on average, compared with 13.61% for the strongest listed baseline, VLA-Adapter-Pro. Each method was tested on 1,800 trials, and the baselines were fine-tuned on the same benchmark data. The revised paper also reports tests across 16 physical-robot tasks using Franka and AgileX PiPER arms; each method-task pair had 60 trials. The authors report stronger alignment in the real-world trials as well, but do not give a pooled success rate across the 16 tasks.
The result has limits. In simulation, DynamicVLA succeeded in 26.5% of disturbance-robustness trials, well below its results for visual and motion generalization. The authors say they omitted real-robot disturbance tests because pushes and contact conditions were difficult to reproduce consistently. The findings are controlled experiments reported by the paper’s authors, not an independent replication or evidence of deployment.
Sources
- DynamicVLA: A Vision-Language-Action Model for Dynamic Object Manipulation, revised October 1, 2026 (arXiv)
- Authors’ project page
- Illustrative image: Jo Teichmann / KUKA Roboter GmbH, public domain, Wikimedia Commons. The pictured robot is not the system evaluated in the paper.
