World-action models (WAMs) are meant to be generalist robot policies that jointly predict how the world will evolve and what actions to take. But predicting a plausible future is not enough, argue the authors of a new arXiv preprint: in long-horizon manipulation, similar observation-action segments recur across different tasks and stages, leading the model to produce a locally believable rollout that veers toward the wrong part of the task. The researchers call this failure mode Stage Confusion.
To address it, they propose TaskWAM, a closed-loop, stage-aware framework. TaskWAM adds Mode Tokens to represent fine-grained task requirements and State Tokens to track execution progress, together forming a structured task state that conditions the world-action generation. The authors also introduce StageBench, a multi-stage manipulation benchmark with parameterized stage variations for controlled evaluation.
Experiments on simulation benchmarks and real-world tasks show that TaskWAM reduces stage confusion and achieves competitive performance with state-of-the-art methods. A notable claimed advantage is plug-and-play transfer across heterogeneous WAM architectures and robot embodiments, including single-arm robots, dual-arm grippers, dexterous hands, and humanoid robots. As a single-source report, the paper's claims have not yet been independently replicated, but the benchmark and framework are offered as a step toward more reliable long-horizon embodied policies.