Cooking egg cake
1:29Tool use and coordinated bimanual manipulation across four stages.
VelaScaling Vision-Language-Action Models
with Adaptive Action Curve Parametrization
1Fudan University2Shanghai Innovation Institute3The Chinese University of Hong Kong4NovaXBot
Potato shredding
1:13Sustained contact and precise coordination between the potato and grater.
The idea
From actions to trajectories.
Vela predicts a compact curve and its temporal span.
The same control-point budget covers long, smooth motions or shorter, finely resolved movements.
01 Compact geometry
A compact set of spline control points defines a continuous action trajectory.
02 Adaptive horizon
A motion-dependent horizon adapts the temporal span to smooth or precise movements.
03 Scalable pretraining
A shared action interface supports learning across heterogeneous robot embodiments.
Pretrained in trajectory space on approximately 40,000 hours of public and private robot data.
Evaluation
Across tasks. Across control demands.
From real-world bimanual manipulation to diverse simulation benchmarks.
Real-world manipulation
Mean subtask success rate · four stages · 10 trials per stage
Cooking egg cake
67.5%
Potato shredding
87.5%
Task progressionCooking
egg cake
Task progressionPotato
shredding
Simulation benchmarks
LIBERO-X and EBench
Vela retains the same architecture, trainable parameter count, and input modalities as π0.5, changing the output modeling space to action curves. Benchmark post-training follows the official π0.5 recipes.
LIBERO-X
Average success rate
45.3%
EBench
Overall success rate
49.7%
Gains are absolute percentage-point improvements over π0.5. Results follow the manuscript’s evaluation protocols.
Benchmark details
LIBERO-X
Success rate (%)| Model | L1 | L2 | L3 | L4 | L5 | Avg. |
|---|---|---|---|---|---|---|
| OpenVLA-OFT | 29.0 | 17.6 | 8.8 | 6.4 | 4.2 | 13.2 |
| π0 | 29.4 | 21.9 | 11.0 | 7.6 | 5.1 | 15.0 |
| X-VLA | 30.1 | 22.6 | 10.3 | 6.0 | 4.1 | 14.6 |
| OpenWAM-α | 35.9 | 28.2 | 21.6 | 13.0 | 10.0 | 21.7 |
| GR00T N1.5 | 43.3 | 32.9 | 18.7 | 13.3 | 9.7 | 23.6 |
| τ0-VLA | 46.5 | 34.2 | 20.5 | 13.0 | 10.5 | 24.9 |
| Xiaomi-R0 | 71.4 | 49.6 | 33.7 | 22.1 | 17.2 | 38.8 |
| π0.5 | 65.2 | 53.2 | 36.0 | 24.1 | 18.0 | 39.3 |
| π0.5-Vela | 66.1 | 53.5 | 40.7 | 27.3 | 22.5 | 42.0 |
| Vela | 67.2 | 55.9 | 44.9 | 31.5 | 27.1 | 45.3 |
EBench
Success rate (SR) and task-progress Score (%)| Model | Tabletop | Pick & place | Long horizon | Overall | ||||
|---|---|---|---|---|---|---|---|---|
| SR | Score | SR | Score | SR | Score | SR | Score | |
| X-VLA | 8.6 | 24 | 50.0 | 54 | 6.2 | 25 | 23.7 | 36 |
| InternVLA-A1 | 4.3 | 11 | 43.0 | 47 | 17.9 | 46 | 23.9 | 36 |
| FastWAM | 2.9 | 13 | 49.5 | 53 | 16.9 | 38 | 25.6 | 37 |
| Cosmos3-Edge | 18.6 | 32 | 41.5 | 44 | 24.1 | 49 | 29.3 | 42 |
| GigaBrain-0.7 | 37.9 | 59 | 42.0 | 45 | 20.0 | 37 | 33.3 | 46 |
| π0 | 32.9 | 49 | 41.5 | 45 | 25.7 | 49 | 33.7 | 47 |
| InternVLA-A1.5 | 12.1 | 22 | 50.0 | 53 | 33.7 | 59 | 34.2 | 46 |
| Qwen-RobotManip | 50.0 | 70 | 56.5 | 60 | 29.9 | 55 | 45.6 | 60 |
| π0.5 | 29.3 | 45 | 56.5 | 61 | 34.1 | 55 | 41.4 | 54 |
| π0.5-Vela | 34.3 | 54 | 59.0 | 63 | 28.5 | 55 | 42.0 | 58 |
| Vela | 44.3 | 64 | 63.0 | 67 | 39.1 | 65 | 49.7 | 66 |
π0.5-Vela is initialized from the π0.5 checkpoint and post-trained in trajectory space; Vela is pretrained directly in trajectory space.
EBench · capability profile
Task-progress Score (%) on EBench.
EBench · comparison across capabilities
Citation
@misc{li2026vela,
title = {Vela: Scaling Vision-Language-Action Models with Adaptive Action Curve Parametrization},
author = {Yifan Li and Jiaxu Wang and Dongming Wu and Yicheng Jiang and Ryan Ji and Xiangyu Yue and Yanwei Fu},
year = {2026}
}