Long-horizon robot manipulation remains a formidable challenge, requiring robots to both execute individual skills reliably and sequence them coherently over extended tasks. Most hierarchical vision-language-action (VLA) models make each decision with a single forward pass, leaving no room to allocate extra computation to difficult or consequential choices.
To address this, researchers at Hugging Face introduce τ₀-VLA, a hierarchical robot foundation model that treats high-level subtask generation as a compute-scalable inference problem. At each step, the high-level policy uses execution memory to propose a subtask and, when uncertain, searches over alternatives before committing. A low-level policy then executes the chosen subtask across multiple robot embodiments.
The model is trained on 40,115 hours of heterogeneous real-world data with multimodal co-training. The key innovation is world-model-guided test-time computation: the high-level policy can simulate the visual consequences of candidate subtasks using a world model, enabling it to compare possible futures before acting.
Experiments show that selective test-time computation improves next-subtask prediction accuracy by 15–24 percentage points across in-domain and distribution-shifted settings. In real-world tasks with 13–25 ordered steps and episodes lasting up to 12 minutes, hierarchical planning raises average closed-loop success from 27.5% to 45.0% using the same low-level policy.
The team has released the official code and pretrained low-level VLA checkpoint, with the high-level policy to follow. This work opens new avenues for scaling computation in robot learning, making long-horizon tasks more tractable.