τ(0)-VLA uses 40,115 hours of robot data to tackle long-horizon tasks

机器之心 · wechat · 2026-07-27

A long WeChat article breaks down τ(0)-VLA, a new embodied AI model aimed at long-horizon robotic tasks. The key idea is to split planning and execution: a high-level “slow thinking” policy handles task decomposition and decision making, while a low-level “fast execution” policy handles real-time control.

What the model adds

Introduces test-time computation and a world-model-guided planning loop for embodied decision making
Uses a four-part high-level stack: proposal model, world model, value model, and reflection model
Builds a beam-search-like mechanism over sub-tasks to reason about chain effects in the physical world

Training and scale

Trained on 40,115 hours of real robot interaction data
Includes 20,000+ hours of on-robot data across multiple robot forms and public datasets
Supports multiple embodiments via a unified 40-dimensional action space and masking for different robots

Reported results

On AGIBOT G1, hierarchical planning raises average success rate from 27.5% to 45.0% and progress from 80.10% to 87.85%
The model also shows strong direct-execution performance across several manipulation tasks, with some tasks reaching 10/10 success

The article argues that embodied AI is moving from short demo actions toward long-horizon real-world completion, and that “thinking before acting” may be the key bottleneck to solve.

Original post →

τ(0)-VLA uses 40,115 hours of robot data to tackle long-horizon tasks

More from Embodied