Vela uses spline-based parametrization for continuous vision-language-action trajectories
October 3, 2026
Vela replaces fixed-rate action chunks with a compact spline-based representation that adapts temporal support based on motion. This approach allows a fixed output budget to scale across varying prediction horizons and maintains smoothness for contact-rich manipulation tasks.
HOW THIS AFFECTS YOU
●
builderThis enables more precise control in contact-rich manipulation tasks without increasing the model's output budget.
●
researcherYou can move beyond fixed-rate action chunks to model continuous robot trajectories more efficiently.