Vela: a VLA that predicts smooth trajectories instead of action chunks

In short: Most vision-language-action (VLA) models output a fixed number of future actions at fixed time steps, called an "action chunk." Vela instead predicts the control points of a smooth curve plus how long that curve should last. The same output budget can then cover a long, smooth reach or a short, precise contact move.

The problem it fixes

With action chunks, neighboring actions are almost identical, so capacity is wasted, and there is a trade-off: a longer horizon means coarser steps, which hurts precise, contact-rich manipulation.

How Vela works

• Predicts 12 cubic B-spline control points and a motion-dependent horizon (16–40 steps).
• During training, each demonstration gets the longest horizon the spline can represent within a noise-aware tolerance.
• A shared 26-dimension action space lets it train across many robot types.
• Built on the π0.5 architecture (3B PaliGemma backbone plus a 300M action expert), pretrained on about 20,000 hours of public robot data from 60 datasets plus about 20,000 hours of private data.

Results reported by the authors

Benchmark	π0.5	Vela
LIBERO-X average success	39.3%	45.3%
EBench overall success	41.4%	49.7%
Real egg-cake cooking (mean subtask success)	37.5%	67.5%
Real potato shredding (mean subtask success)	62.5%	87.5%

The gains on LIBERO-X were largest at the harder distribution-shift levels. Simply fitting a spline after a normal model's output did not recover the gains.

Why it matters

How a robot model represents actions is as important as how big it is. Vela suggests continuous trajectories could become the default output format for future robot foundation models. These are the authors' own numbers on a preprint and have not yet been independently replicated.