Event

Proposes TaskWAM framework and StageBench benchmark

1 reports · 1 sources

Researchers including Zhuohao Li propose TaskWAM—a closed-loop, stage-aware world-action model framework that uses Mode Tokens to represent fine-grained task requirements and State Tokens to track execution progress, forming a structured task state to guide the world-action model in generating futures that match the current task, thereby addressing the Stage Confusion phenomenon in long-horizon manipulation. The paper also proposes StageBench, a multi-stage manipulation benchmark with parameterized stage variations and automated evaluation. The authors state that experiments on simulation benchmarks and real-world tasks show that TaskWAM effectively reduces confusion, achieves performance comparable to state-of-the-art methods, and can be transferred plug-and-play to different world-action model architectures and robot embodiments, including single-arm robots, dual-arm grippers, dexterous hands, and humanoid robots.