τ₀-VLA
Summary
τ₀-VLA is a hierarchical robot foundation model that maintains correctable execution memory, uses world-model-guided test-time search to choose language subtasks, and hands the selected subtask to a low-level flow-matching VLA over continuous robot control inputs.
Role In The Wiki
τ₀-VLA is the clearest current local example of adaptive compute at the subtask boundary. It sits between Gemini Robotics 1.5’s high-level reasoning handoff and π0.7’s visual-subgoal conditioning: τ₀-VLA predicts candidate terminal images before commitment so imagined consequences can change the selected subtask.
Evidence
Official Artifacts
The captured code and approximately 3B-parameter checkpoint expose the low-level VLA stack. The high-level proposal/world/value/reflection/search and execution-memory components are not visible in the public artifact surface, so the complete hierarchy is not currently reproducible.
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in tau0-vla-2026 rather than duplicating verdict rows here.
At the entity level, τ₀-VLA is a robotics analogue for a digital-world robot that maintains execution state, proposes typed interventions at sparse decision boundaries, predicts their consequences, allocates more compute to uncertain decisions, and sends the selected intervention to a faster executor. The source page records why terminal-image prediction and open-ended language subtasks are not yet a numeric action-conditioned time-series world model.