World models without compositional reuse
Robotics research has settled into two contrasting paradigms. World action models deliver visual foresight and future prediction for manipulation, yet they typically omit explicit mechanisms for reusing skills across tasks. Agentic robotics, by contrast, blends code-based planning with reusable skills but often abandons predictive grounding in physical dynamics. The paper Agentic World Action Models bridges this divide.
Traditional world models operate as end-to-end predictors of future pixels given a sequence of actions. They excel at short-horizon foresight but struggle when asked to reason about longer sequences of behavior. The underlying assumption—that a single policy can implicitly capture all necessary skills—breaks down as task complexity grows. Meanwhile, agentic systems that embed programmatic control logic gain compositional power but lose the tight coupling between predicted visual states and executed motor commands.
The AW-0 architecture
AW-0 introduces a unified mixture-of-transformers architecture that jointly models three coupled generative streams: code, future video, and robot actions. Rather than treating these as separate modules, the framework interweaves them so that each stream conditions on the others. A code expert generates parameterized skill routines and control logic; video and action experts respectively produce future visual states and low-level motor commands. This coupling means that a predicted action influences the next video state, which in turn informs the next code generation, creating a loop of mutual refinement.
Code serves two complementary roles. First, it structures multi-step prediction and control within the world model, determining which skill to activate and how to parameterize it at each timestep. Second, and equally important, the generated code becomes a reusable program that the agent can execute directly after planning finishes. This dual function means that the same signal that guides prediction also yields executable skills for downstream agentic behavior.
Automated skill curation
A key contribution of the paper is an automated curation pipeline that labels robot and human video trajectories with aligned programs. The pipeline scans execution traces, extracts recurring routine patterns, and consolidates them into reusable skill libraries. By grounding these routines in both demonstration data and simulated rollouts, the system builds a library of behaviors that can be composed without re-training the entire model. The result is a growing repository of skills that the agent can select from during execution.
Experimental results
Evaluations on RoboTwin 2.0 and physical bimanual manipulation tasks demonstrate that AW-0 outperforms VLA and WAM baselines on several dimensions. Task success rates improve, and the framework shows measurable gains in language following, semantic reasoning, and robustness. These results suggest that the joint modeling of code, video, and actions produces more coherent behavior than approaches that treat planning, prediction, and execution as separate components.
Limitations and trade-offs
The mixture-of-transformers design introduces additional compute overhead compared to simpler predictive models. The curation pipeline, while effective at surfacing recurring routines, depends on the quality and diversity of the training trajectories; sparse or noisy demonstrations may yield less useful skill libraries. Additionally, the three-stream coupling means that errors in one stream can propagate through the others, requiring careful training dynamics to stabilize.
Practical takeaways
For developers working with robotic manipulation systems, AW-0 offers a concrete pathway to integrate code-based planning without sacrificing visual foresight. The framework's architecture can be adopted by replacing a single-stream world model with a tripartite mixture that jointly learns code, video, and action generation. The curated skill library provides an immediate benefit: instead of learning every behavior from scratch, systems can draw from a growing pool of proven routines. Future work may explore hierarchical skill composition and real-time adaptation of the code expert to new domains.