τ_0-VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation
The model spends extra inference compute on hard high-level robot decisions before choosing the next subtask.
The paper presents a hierarchical VLA system where a high-level policy can search over subtask alternatives using a world model, instead of making every choice in one pass. A low-level policy then carries out the selected subtask across multiple robot embodiments. The authors say it was trained on 40,115 hours of heterogeneous real-world data. In their reported tests, more test-time computation improved next-subtask prediction and raised closed-loop success on long-horizon manipulation tasks. HF Daily Papers' note
The paper presents a hierarchical VLA system where a high-level policy can search over subtask alternatives using a world model, instead of making every choice in one pass. A low-level policy then carries out the selected subtask across multiple robot embodiments. The authors say it was trained on 40,115 hours of heterogeneous real-world data. In their reported tests, more test-time computation improved next-subtask prediction and raised closed-loop success on long-horizon manipulation tasks. HF Daily Papers' note
score 5