A new preprint on arXiv challenges the standard approach to vision-language-action (VLA) models for robotics. The authors argue that robotic manipulation is not merely a matter of converting observations into actions, but rather of achieving task-specified changes in the physical world. They propose that existing methods, which directly learn observation-to-action mappings, may miss this fundamental goal.

The paper introduces a method called Vision-Language-Delta-Action, which instead learns compact deltas—small, actionable changes—as a transition interface between perception and action. By focusing on the difference between current and desired states, the approach aims to make manipulation more grounded in the actual physical outcome of the task.

Because this is a single source, there are no conflicting findings to compare. The abstract is truncated, so details on experimental results or implementation are not available from the provided text. The core idea, however, offers a conceptual shift in how VLA models might be structured for robotic control.