Hugging Face Releases Vela: Revolutionizing Robot Behavior Modeling with Continuous Trajectory Representations
Summary:Hugging Face has released Vela, a vision-language-action (VLA) foundation model for robot behavior modeling. Vela introduces a spline-based compact action representation, motion-dependent temporal support, and a shared action interface for heterogeneous embodiments, allowing for adaptive temporal resolution across motions within a fixed output budget. Unlike traditional fixed-rate action chunk representations, Vela more efficiently captures the continuity and smoothness of robot behavior and dem
Background and Challenges
Traditional vision-language-action (VLA) models typically represent future actions as fixed-rate action chunks, which ties temporal resolution and prediction horizon to a fixed output budget. This pointwise representation wastes capacity on highly correlated neighboring actions and requires implicitly learning temporal continuity and smoothness. Additionally, it forces a trade-off between long-horizon coverage and the local precision required for contact-rich manipulation, making it difficult to meet the demands of complex tasks.
Core Innovations of Vela
Vela addresses these issues through the following innovations:
-
Spline-Based Action Representation: Vela uses a compact spline-based representation for action trajectories instead of fixed-rate action chunks. This approach more efficiently captures the continuity and smoothness of actions.
-
Motion-Dependent Temporal Support: Vela introduces a motion-dependent temporal support mechanism, allowing temporal resolution to adapt based on the dynamic characteristics of the action, thus enabling more flexible time allocation within a fixed output budget.
-
Shared Action Interface for Heterogeneous Embodiments: Vela features a shared action interface that can adapt to different types of robotic entities, ensuring the model's versatility across various hardware platforms.
Experiments and Results
Vela was pretrained on multiple large-scale multi-embodiment robot datasets and evaluated on LIBERO-X, EBench, and two real-world long-horizon tasks: egg-cake cooking and potato shredding. The results demonstrate Vela's effectiveness in both simulated and physical operations, validating the advantages of its continuous action representation.
Engineering Trade-offs and Practical Applications
While Vela's continuous action representation improves temporal resolution and action smoothness, it also introduces higher computational costs and a more complex training process. Additionally, the parameterization choices for splines and the motion-dependent temporal support mechanism require careful tuning based on specific tasks to ensure the model's generalization across different scenarios.
Developer Recommendations
For developers looking to adopt Vela, the following recommendations are suggested:
-
Data Preparation: Ensure high-quality multi-embodiment robot data is provided to fully leverage Vela's pretraining advantages.
-
Model Fine-Tuning: Fine-tune Vela according to the specific application scenario, particularly in terms of temporal resolution and action representation.
-
Hardware Acceleration: Utilize GPU or TPU hardware acceleration tools to address Vela's high computational demands.
Future Outlook
The release of Vela lays the foundation for future robot behavior modeling. Further exploration of its applications in multimodal interaction, complex task planning, and dynamic environment adaptation is anticipated.
— END —Source: Hugging Face Daily Papers (2026-10-04)
Tags: #Hugging Face #Vela #VLA Model #Robot Behavior Modeling #Continuous Trajectory Representation
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments