WLA-0 Unifies Robot Physics and Language Reasoning, Hits 92.94% on Manipulation Benchmark
WLA-0 merges world modeling, language reasoning, and robot action into one 2B-parameter model that hits 92.94% success on RoboTwin2.0 at 40ms per inference
Takeaways
- New model class: WLA (World-Language-Action) unifies world modeling, language reasoning, and robot action in one autoregressive Transformer.
- Dual expert design: A World Expert predicts future visual states; an Action Expert uses those dynamics to generate more precise robot actions.
- SOTA results: WLA-0 (2B params) hits 92.94% on RoboTwin2.0 Clean and 56.5% on the memory-dependent RMBench benchmark.
- 40ms inference: World modeling can be disabled at runtime via meta-queries, enabling fast deployment on an NVIDIA RTX 5090.
- Test-time scaling: Activating the World Expert at inference lets the model imagine and score candidate futures before acting -- trading latency for accuracy.
- Action-free learning: WLA-0 can learn novel tasks from cross-embodiment robot videos without any action annotations.
The robotics AI field has been split between two camps: Vision-Language-Action (VLA) models that are great at following instructions but clumsy with physics, and World-Action Models (WAMs) that understand physical dynamics but struggle with high-level task reasoning. A new paper proposes World-Language-Action (WLA) models as a unified class that does both -- and backs it up with state-of-the-art numbers on two major benchmarks.
Two worlds, one gap
VLA models integrate visual perception, natural language understanding, and robotic action into a unified framework, enabling machines to perform complex tasks in dynamic, unstructured environments. But they have a fundamental blind spot. These models still face bottlenecks in coping with the high diversity and uncertainty of the physical world -- they are more like "imitating" patterns in training data, lacking the foresight of action consequences and the understanding of physical logic.
On the other side, WAMs address this by training on video. A world action model (WAM) jointly predicts future world states and robot actions using video pretraining. By training on large-scale video data -- including internet video and egocentric human footage -- a WAM develops an internalized model of physics, motion, and interaction that transfers broadly to new tasks and environments. The catch? WAMs predict future visual states to understand dynamics, but often struggle with high-level semantic reasoning. They know how things move, but not necessarily why in a complex, task-oriented sense.
WLA is the attempt to close that gap. It takes the world modeling interface from WAMs and the language reasoning from VLAs, and fuses them into a single autoregressive Transformer.

The architecture: two experts, one backbone
At its core, WLA predicts the "next state" as two complementary representations. First, it predicts a high-level textual intention, like a subtask, which provides abstract semantic guidance. Second, it predicts low-level physical dynamics, which captures the fine-grained visual transition needed for precise motor control. This dual prediction is handled by two specialized modules sitting on top of a shared autoregressive Transformer backbone: