Vela: Scaling Vision-Language-Action Models with Adaptive Action Curve Parametrization

Vela. Scaling Vision-Language-Action Models with Adaptive Action Curve Parametrization. By Yifan Li and six others, from Fudan University, the Shanghai Innovation Institute, and two other institutions. Published on the fourth of October. No code repository is linked to it.

Most vision-language-action models describe future motion as action chunks, emitted at a fixed rate. Each chunk is a row of individual actions, evenly spaced in time. That ties two things to a fixed output budget at once: the temporal resolution, meaning how finely time is sliced, and the prediction horizon, meaning how far ahead the model looks.

This pointwise representation has three costs. It spends capacity on neighbouring actions that are highly correlated, and so mostly repeat each other. It leaves temporal continuity and smoothness to be learned implicitly, rather than building them in. And it forces a trade. With a fixed budget, covering a long horizon means coarser steps, while the contact-rich parts of manipulation need fine local precision. A model cannot have both from the same row of points.

Vela is a vision-language-action foundation model that changes the unit of output. It represents future robot behaviour as continuous trajectories, curves through time, instead of sequences of points.

It combines three pieces. A compact action representation based on splines. Temporal support that depends on the motion being made. And a shared action interface for heterogeneous embodiments, so that robots with different bodies speak through one format. Together, these let a fixed output budget adapt its temporal resolution from one motion to the next, rather than spending it the same way every time.

Vela is pretrained on large-scale robot data drawn from multiple embodiments. It is then evaluated in simulation on LIBERO-X and EBench, and on two real-world long-horizon tasks: cooking an egg cake, and shredding potato. The authors describe the results across simulation and physical manipulation as promising.

The authors read these results as pointing to the potential of continuous action representations as a foundation for future embodied foundation models. The shift is easy to state: a curve that can spend detail where the motion demands it, in place of a row of evenly spaced points.