τ₀-VLA hits 45% success with world-model test-time compute
TL;DR
- τ₀-VLA is trained on 40,115 hours of heterogeneous real-world robot data with a Qwen3.5 backbone and a Mixture-of-Transformers action expert.
- Hierarchical planning reaches 45.0% average success on long-horizon manipulation, up from 27.5% for flat execution; π0.5 baseline scores 22.5%.
- World-model-guided search at plan time lifts a distribution-shifted book-organization task to 74.0% from 50.0% under a single-shot planner.
The τ₀-VLA robot foundation model, released on arxiv, was trained on 40,115 hours of heterogeneous real-world robot data. Its hierarchical system, in which a high-level planner proposes subtasks and a low-level policy executes them, reaches 45.0% average success on long-horizon manipulation. The flat model, without hierarchy, gets 27.5%.
The move that lifts the number is a search step at planning time. "It proposes subtasks, predicts their visual outcomes, and compares the resulting branches before commitment," the project page reports. On a book-organization task under distribution shift, this test-time computation pushes success to 74.0% against 50.0% for the same system planning once. A milk-tea preparation task goes from 5/10 to 7/10.
"Confidence-based routing enables selective test-time computation by allocating additional reasoning only when uncertain," the authors write, framing the choice as spending compute on the hard steps and skipping it on the easy ones. Selective compute gains 15 to 24 percentage points on next-subtask accuracy; a revisable memory adds another 11.0.
The comparison bar is low. GR00T N1.7 scores 2.5% on the same long-horizon suite, LingBot-VLA 0.0%, π0.5 22.5%. Neither the abstract nor the project page publishes per-task variance or the size of the eval batch beyond the 5/10 and 7/10 tallies.
Originally reported by paper
Read the original article →Original headline: τ₀-VLA Trains Hierarchical Robot Policy on 40K Hours of Real Data, World-Model Search Lifts Success 63%