IllustrationNo official image on file. Showing a schematic category illustration.
Unified VLA model combining manipulation, navigation and trajectory prediction in one framework (research release).
Key specs
| Architecture | Qwen3.5-4B + 1.15B DiT decoder |
|---|
Technical details
Qwen-VLA is Alibaba Tongyi/Qwen's unified VLA that extends the Qwen vision-language stack to continuous action and trajectory generation. It couples a Qwen3.5-4B backbone (early-fusion ViT tokens, hybrid gated-linear/GQA attention) with a 1.15B DiT flow-matching action decoder, and casts manipulation, navigation and trajectory prediction into one shared prediction space. Embodiment-aware prompt conditioning prepends a textual robot description to select embodiment and control convention, avoiding per-platform heads. The team reports that a single generalist matches or outperforms per-benchmark specialists across simulation and real-world evaluations.
| Architecture | Qwen3.5-4B natively multimodal vision-language backbone plus a 1.15B single-stream DiT flow-matching action decoder; embodiment-aware prompt conditioning (robot-specific text prepended) instead of per-platform output heads; unified action-and-trajectory prediction space. |
|---|---|
| Params | ~5.15B total (Qwen3.5-4B backbone + 1.15B DiT action decoder) |
| Inputs outputs | Inputs: images/video, language instructions, embodiment description prompt; outputs: continuous action chunks, navigation waypoints and trajectories. |
| Training data | Large-scale joint pretraining on robot manipulation trajectories, human egocentric demonstrations, synthetic simulation data, vision-and-language navigation data, trajectory-centric supervision and auxiliary vision-language data; plus text-to-action DiT pretraining with frozen VLM. |
| Embodiments supported | Single-arm, dual-arm and humanoid robots; Tongyi states one model deploys to more than 10 distinct platforms without architecture changes. |
| How to try | github.com/…/Qwen-VLA |
| Sources |
