arxiv:2608.16885

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

Published on Aug 17

· Submitted by

jrryzh(SII) on Aug 21

Shanghai Innovation Institute

Upvote

Authors:

Abstract

A hierarchical vision-language-action model improves long-horizon robot manipulation by using world-model-guided test-time search to scale computation for high-level subtask decisions.

Generated by thinkingmachines/Inkling-Small

Long-horizon robot manipulation requires a robot to both execute individual skills reliably and sequence them coherently over extended tasks. Most hierarchical vision-language-action (VLA) models make each such decision with a single forward pass, leaving no mechanism to allocate additional computation to difficult or consequential choices. We introduce τ_0-VLA, a hierarchical robot foundation model that formulates high-level subtask generation as a compute-scalable inference problem through world-model-guided test-time computation. At each inference step, the high-level policy uses execution memory to generate a subtask and, when needed, searches over alternatives before committing to its output. A low-level policy then executes the generated subtask across multiple robot embodiments. The policy is trained on 40,115 hours of heterogeneous real-world data with multimodal co-training. Across in-domain and distribution-shifted settings, allocating additional test-time computation substantially improves next-subtask prediction accuracy, and these gains translate into higher closed-loop success on long-horizon robot manipulation tasks.

View arXiv page View PDF Project page GitHub 518 Add to collection

Community

J3rr1

Paper author Paper submitter 1 day ago

🤖 What if a robot could compare possible futures before deciding what to do next?

We introduce τ₀-VLA, a hierarchical robot foundation model for long-horizon manipulation. Its high-level policy maintains execution memory and, when a decision is uncertain, allocates additional test-time computation to propose candidate subtasks, predict their visual consequences with a world model, and compare alternatives before committing. A generalist low-level VLA then executes the selected subtask across robot embodiments.

Highlights:

The low-level policy is trained on 40,115 hours of heterogeneous real-world robot data with multimodal co-training.
Selective test-time computation improves next-subtask prediction accuracy by 15–24 percentage points across in-domain and distribution-shifted settings.
We evaluate real-world manipulation tasks containing 13–25 ordered steps, with episodes lasting up to 12 minutes.
Using the same low-level policy, hierarchical planning improves average closed-loop success from 27.5% to 45.0% across four long-horizon tasks.
We release the official code and pretrained low-level VLA checkpoint, with high-level policy on the way.

🌐 Project page
💻 Code
🤗 Model checkpoint

Questions and feedback are very welcome!