Shanghai Innovation InstituteChina2026

τ₀-VLA

An AI model: a vision-language-action model (VLA) takes camera images and an instruction and turns them into movements for a robot. This page describes the software, not a robot.

RobotsG1by Agibot B

Key figures as stated by the developer

5 of 6 stated · 1 gap

Model type ⁠A vision-language-action
Controls ⁠A arms and hands
Inputs ⁠A camera language robot state
Access ⁠A open download
Parameters not stated
Base model ⁠A Qwen3.5-2B

How to read the figures

50HzA: The letter in the grey square after a figure points to the source the figure came from. Pages with several sources carry several letters.

What the data labels mean

Versions and notes in the key figures (6)

Model type

Model card, Model type: "Pretrained low-level vision-language-action policy".

Controls

Model card, Output: "A normalized `30 × 40` continuous action chunk" "The model uses a unified 40-dimensional state/action layout" with end-effector, gripper, arm-joint, waist and chassis-velocity slots.

Inputs

Model card, Inputs: "Multi-view RGB observations, language instruction or subtask, robot metadata, state, and masks". RGB observations = camera, state = robot state.

Access

Model card, Post-training: "hf download sii-research/tau-0-vla".

Parameters

Note Repository README: "The low-level policy combines a Qwen3.5 vision-language backbone with a Mixture-of-Transformers action expert trained through conditional flow matching."

Base model

Model card header: "base_model: Qwen/Qwen3.5-2B".

Full specifications

10 of 13 fields stated

The letter points to the source at the foot of the page. Fields without a figure are dashed: the developer does not state them — that is not the same as zero, and not the same as no.
Field	Value	Source and caveats
Input and output
Model type	`vision-language-action`	A Model card, Model type: "Pretrained low-level vision-language-action policy".
Inputs	`camera` `language` `robot state`	A Model card, Inputs: "Multi-view RGB observations, language instruction or subtask, robot metadata, state, and masks". RGB observations = camera, state = robot state.
Controls	`arms and hands`	A Model card, Output: "A normalized `30 × 40` continuous action chunk" "The model uses a unified 40-dimensional state/action layout" with end-effector, gripper, arm-joint, waist and chassis-velocity slots.
Degrees of freedom controlled	not stated	Note Model card, the one robot route in the public code: "Dual 7-DoF arms and grippers". ⁠A
Size and speed
Parameters	not stated	Note Repository README: "The low-level policy combines a Qwen3.5 vision-language backbone with a Mixture-of-Transformers action expert trained through conditional flow matching." ⁠C
Control rate	approximately 30Hz	B Action execution layer of the low-level policy, paper appendix on serving: it "keeps executing the current cached subtask at the control rate" of about 30 Hz while the high-level policy refreshes the subtask.
Where it runs and what it is built on
Base model	Qwen3.5-2B	A Model card header: "base_model: Qwen/Qwen3.5-2B".
Openness
Access	open download	A Model card, Post-training: "hf download sii-research/tau-0-vla".
License	Apache-2.0	A Model card header: "license: apache-2.0".
Free or paid	fully open source	A Weights and code, model card header: "license: apache-2.0".
The developer’s own claims
Learns new tasks	needs retraining	A Developer claim: "It is intended as an initialization for embodiment- and task-specific post-training." "The checkpoint is not demonstrated as a zero-shot controller for arbitrary robots or tasks."
Works across robots	yes	D Developer claim, abstract: "A low-level policy then executes the generated subtask across multiple robot embodiments." Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count.

Not stated: On robot or cloud

Notes

Shanghai Innovation Institute calls τ₀-VLA a hierarchical robot foundation model for long-horizon manipulation.
The τ₀-VLA project website is dated 27 July 2026.
The model card describes only the low-level policy checkpoint; the full system adds high-level models for test-time search.
The paper lists Shanghai Innovation Institute as the first affiliation.

Sources

A huggingface.co/sii-research/tau-0-vla · Retrieved 2026-10-15
B arxiv.org/html/2608.16885v1 · Retrieved 2026-10-15
C github.com/sii-research/tau-0-vla · Retrieved 2026-10-15
D arxiv.org/abs/2608.16885 · Retrieved 2026-10-15

One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.