τ_0-VLA: A Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation

Summary

Long-horizon robot manipulation requires both reliable execution of individual skills and coherent sequencing over extended tasks. Most hierarchical vision-language-action (VLA) models make decisions with a single forward pass, lacking a mechanism to allocate additional computation to difficult choices. We introduce τ_0-VLA, a hierarchical robot foundation model that frames high-level subtask generation as a compute-scalable inference problem via world-model-guided test-time computation. At each step, the high-level policy uses execution memory to generate a subtask, searching over alternatives when needed, while a low-level policy executes across embodiments. Trained on 40,115 hours of heterogeneous real-world data with multimodal co-training, the model shows that additional test-time computation substantially improves next-subtask prediction accuracy in both in-domain and distribution-shifted settings, translating to higher closed-loop success on long-horizon tasks.