Top AI Product

Qwen-VLA extends Qwen’s vision-language stack past perception into actual robot control — adding a DiT-based action decoder that turns understanding into continuous actions and trajectories. The goal is one model that handles many robot bodies and tasks, not a separate VLA per platform.

## Cross-embodiment by prompt

The hard part of generalist robot models is that every robot has different actuators, control conventions, and quirks. Qwen-VLA uses embodiment-aware prompt conditioning — a textual description specifies the current robot and its control convention, and the model adapts accordingly. Training joins three data sources: real robot manipulation trajectories, human egocentric demonstrations, and synthetic simulation data, all under one pretraining recipe.

## Benchmark sweep

The reported results span the standard VLA gauntlet: 97.9% on LIBERO, 73.7% on Simpler-WidowX, 86.1% and 87.2% on RoboTwin Easy and Hard, 69.0% OSR on R2R navigation, 59.6% SR on RxR, 76.9% average out-of-distribution success on real-world ALOHA bimanual tasks, and 26.6% zero-shot on DOMINO’s dynamic manipulation. That’s strong simultaneous coverage across navigation, manipulation, and dynamic tasks — usually where generalists drop off.

## Why it matters

Robot foundation models have been heavily fragmented by embodiment. A single VLM lineage that scales into action, with prompt-level embodiment switching, points at a cleaner road to general-purpose robot policies — and at a future where the same VLM you already use sees, reasons, and now also acts.