BAAIChina2025

UniVLA

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.

Robotsnot stated

Key figures as stated by the developer

6 of 6 stated

Model type ⁠A vision-language-action
Controls ⁠B arms and hands
Inputs ⁠A camera language
Access ⁠C open download
Parameters ⁠B 8,500,000,000
Base model ⁠B Emu3 (first-stage weights)

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

Developer wording, abstract: "we present UniVLA, a unified and native multimodal VLA model that autoregressively models vision, language, and action signals as discrete token sequences".

Controls

Real-robot setup, appendix: "The model outputs a 14-dimensional action vector." "The real-robot is controlled using end-effector (EE) pose."

Inputs

Abstract: "we present UniVLA, a unified and native multimodal VLA model that autoregressively models vision, language, and action signals as discrete token sequences". Real-robot setup, appendix: "using images captured from three perspectives". Vision = camera.

Access

Hugging Face model card: "This repository contains the model weights for the paper".

Parameters

Paper, implementation details: "The model adopts a purely autoregressive Transformer architecture with 8.5 billion parameters".

Base model

Paper: "The model is initialized with pre-trained weights from the first stage of Emu3".

Full specifications

9 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 Developer wording, abstract: "we present UniVLA, a unified and native multimodal VLA model that autoregressively models vision, language, and action signals as discrete token sequences".
Inputs	`camera` `language`	A Abstract: "we present UniVLA, a unified and native multimodal VLA model that autoregressively models vision, language, and action signals as discrete token sequences". Real-robot setup, appendix: "using images captured from three perspectives". Vision = camera.
Controls	`arms and hands`	B Real-robot setup, appendix: "The model outputs a 14-dimensional action vector." "The real-robot is controlled using end-effector (EE) pose."
Degrees of freedom controlled	not stated	Note The paper gives a size for the action vector, not a number of controlled degrees of freedom: "The model outputs a 14-dimensional action vector." Robot: "Real-world setup of the AgileX Cobot Magic dual-arm robot". ⁠B
Size and speed
Parameters	8,500,000,000	B Paper, implementation details: "The model adopts a purely autoregressive Transformer architecture with 8.5 billion parameters".
Where it runs and what it is built on
Base model	Emu3 (first-stage weights)	B Paper: "The model is initialized with pre-trained weights from the first stage of Emu3".
Openness
Access	open download	C Hugging Face model card: "This repository contains the model weights for the paper".
License	Apache-2.0	C Model card header: "license: apache-2.0".
Free or paid	fully open source	C Model card header: "license: apache-2.0".
The developer’s own claims
Learns new tasks	needs retraining	B Developer claim: the model is adapted to downstream tasks by a second training phase, "a fine-tuning phase that interleaves actions into the sequence, enabling policy learning on downstream tasks".
Works across robots	not stated	Note The developer shows applications, not one model on several robot models: "We further demonstrate its broad applicability on real-world ALOHA manipulation and autonomous driving". ⁠A 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: Control rate, On robot or cloud

Notes

The authors describe UniVLA as a unified and native multimodal vision-language-action model.
The UniVLA paper from BAAI was submitted to arXiv on 24 June 2025.
The paper states that part of the work was done during an internship at BAAI.
The Beijing Academy of Artificial Intelligence (BAAI) was founded on the initiative of the Beijing Municipal Science and Technology Commission and the Haidian District government.

Sources

A arxiv.org/abs/2506.19850 · Retrieved 2026-10-15
B arxiv.org/html/2506.19850v1 · Retrieved 2026-10-15
C huggingface.co/Yuqi1997/UniVLA · 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.