← Leaderboard

τ0-VLA

VLA
Shanghai Innovation Institute · Aug 17, 2026

Hierarchical VLA whose high-level policy plans subtasks from execution memory and can spend extra test-time computation searching alternatives with a world model, trained on 40,115 hours of real-robot data. Results are self-reported on the authors’ long-horizon real-robot tasks (45.0% average success for the hierarchical system vs 22.5% for π0.5), not on a public benchmark.

Type
VLA
Parameters
—
Backbone
Qwen3.5-2B + Mixture-of-Transformers action expert
Weights
Open
License
Apache-2.0
Commercial use
Allowed
Released
Aug 17, 2026
Robots
ARX AC One, Franka
SO-101
–
LeRobot
–
Inputs
–
VRAM
–

Benchmark results

0 of 13 benchmarks

No benchmark results recorded yet.

Not reported on 13 benchmarks
Wrong or missing number? Report an error