10 products compared
VLA models, compared side by side.
Compare vision-language-action (VLA) models on Emergence Club by architecture, embodiments, training data, license and availability.
| Product | Company | Launch date | Product type | Architecture | Availability | Price |
|---|---|---|---|---|---|---|
| Gemini Robotics 2 | Google DeepMind | 2026-07-30 | VLA | Not stated | Early-access partners | Not stated |
| Gemini Robotics ER 2 | Google DeepMind | 2026-07-30 | Embodied-reasoning VLM | Not stated | Google AI Studio; private preview on Gemini Enterprise Agent Platform | Not stated |
| Gemini Robotics On-Device 2 | Google DeepMind | 2026-07-30 | On-device VLA | Not stated | Not stated | Not stated |
| Qwen-VLA | Alibaba Tongyi Lab | 2026-05-29 | Not stated | Qwen3.5-4B + 1.15B DiT decoder | Not stated | Not stated |
| LimX FluxVLA Engine | LimX Dynamics | 2026-04-20 | Not stated | Not stated | Not stated | Not stated |
| Xiaomi-Robotics-0 | Xiaomi | February 2026 | Not stated | Qwen3-VL-4B-Instruct VLM + DiT (flow matching) | Not stated | Not stated |
| Figure Helix 02 | Figure AI | 2026-01-27 | Not stated | Adds System 0 learned whole-body controller | Not stated | Not stated |
| NVIDIA Isaac GR00T N1.6 | NVIDIA | 2026-01-05 | Open VLA with full-body control for humanoids | Not stated | Not stated | Not stated |
| X-VLA (LeRobot) | Hugging Face | 2025-12-04 | Not stated | Soft-prompted unified transformer | Not stated | Not stated |
| π*0.6 (pi-star 0.6) | Physical Intelligence | 2025-11-18 | Not stated | Not stated | Not stated | Not stated |
What is a VLA model?
A vision-language-action (VLA) model is an AI model that takes in camera images and natural-language instructions and outputs robot actions.
VLA models are one kind of robot foundation model: large models trained to give robots general skills that carry across tasks, and sometimes across different robot bodies. This page also includes related robot foundation models, such as embodied-reasoning and world-action models, when they are listed under Brain / model / chip.
Which fields are compared?
Each row in the table is one listing on Emergence Club. Each column is one field, taken from the official source.
- Company and country: the lab or company that released it.
- Release date: when the model was announced or released.
- Model type: for example VLA, embodied reasoning, or world model, as the company describes it.
- Architecture: the model design as published, such as backbone and action head.
- Parameters: model size, when stated.
- Embodiments: the robot types the model has been shown to control, such as arms, humanoids, or mobile robots.
- Functions: tasks the model targets, such as manipulation, navigation, bimanual work, or whole-body control.
- Training data: the data the company says it trained on.
- Company-stated results: performance results as the company reports them, labeled as the company's own statements.
- Availability: how it is offered, such as open weights, API, preview, or in-house only.
- License: the license on the model card or repository.
- Open source: whether code or weights are released under an open license.
- Ecosystem: platforms and hardware it runs on, when stated.
How is this comparison made?
This comparison uses only data from official sources and shows each field as the company published it.
- Official sources only. Every value comes from a company website, official announcement, paper, or official repository, and each row links to its sources.
- Blank when not stated. If a company has not published a field, the cell says "Not stated." We do not estimate or fill in values.
- No rankings or scores. Emergence Club does not rank, score, or pick a winner.
- Dates as published. Release dates are shown as the company gave them. Month-only dates are shown as a month.
- Units as published. Values are not converted or normalized unless the source gives them that way.
Where can I find open VLA models?
The Models section of Emergence Club resources lists open robot models with their license and official repository.
Frequently asked questions
What is the difference between a VLA model and a robot foundation model?
A VLA model maps images and language to robot actions. Robot foundation model is the broader term for large models that give robots general skills, and it includes VLA models.
Which VLA model is best?
Emergence Club does not rank or score models. Compare architecture, embodiments, availability, and license, and read each official source.
Are the performance numbers independent?
No. Results are shown as the company published them and are labeled as company-stated. Emergence Club does not run its own benchmarks.
Which VLA models are open source?
The Open source and License columns show which models have code or weights released under an open license, as stated in the official repository or model card.
Why are some fields "Not stated"?
"Not stated" means the company has not published that detail. Emergence Club does not estimate missing values.