OpenDriveLabChina2025

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

5 of 6 stated · 1 gap

Model type ⁠C vision-language-action
Controls ⁠A arms and hands
Inputs ⁠A camera language
Access ⁠B open download
Parameters not stated
Base model ⁠A Prismatic-7B VLM (SigLIP and DINOv2 vision, LLaMA-2 language)

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 propose UniVLA, a new framework for learning cross-embodiment vision-language-action (VLA) policies".

Controls

Paper: "a lightweight decoder with only 10.8M parameters to translate latent actions into executable trajectories".

Inputs

Paper, Fig. 3: the policy "which processes projected visual embeddings and tokenized task instructions as inputs to predict latent action tokens in an auto-regressive manner". Visual embeddings = camera, task instructions = language.

Access

Hugging Face model card: "This is the official checkpoint of our RSS 2025 work".

Parameters

Note The developer gives no total. Paper: "Our policy architecture is founded on the Prismatic-7B Vision-Language Model (VLM)"; for fine-tuning, "the total number of trainable parameters is approximately 123M".

Base model

Paper: "Our policy architecture is founded on the Prismatic-7B Vision-Language Model (VLM)".

Full specifications

11 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`	C Developer wording, abstract: "we propose UniVLA, a new framework for learning cross-embodiment vision-language-action (VLA) policies".
Inputs	`camera` `language`	A Paper, Fig. 3: the policy "which processes projected visual embeddings and tokenized task instructions as inputs to predict latent action tokens in an auto-regressive manner". Visual embeddings = camera, task instructions = language.
Controls	`arms and hands`	A Paper: "a lightweight decoder with only 10.8M parameters to translate latent actions into executable trajectories".
Degrees of freedom controlled	7DoF	A Stated for the real-robot setup: "a Piper arm from AgileX Robotics featuring a 7-DoF action space".
Size and speed
Parameters	not stated	Note The developer gives no total. Paper: "Our policy architecture is founded on the Prismatic-7B Vision-Language Model (VLM)"; for fine-tuning, "the total number of trainable parameters is approximately 123M". ⁠A
Control rate	10Hz	A This is the model's inference rate: the paper reports a "real-time, closed-loop inference frequency of 10Hz on an NVIDIA RTX 4090 GPU".
Where it runs and what it is built on
Base model	Prismatic-7B VLM (SigLIP and DINOv2 vision, LLaMA-2 language)	A Paper: "Our policy architecture is founded on the Prismatic-7B Vision-Language Model (VLM)".
Openness
Access	open download	B Hugging Face model card: "This is the official checkpoint of our RSS 2025 work".
License	Apache-2.0	B Model card header: "license: apache-2.0".
Free or paid	fully open source	B Model card header: "license: apache-2.0".
The developer’s own claims
Learns new tasks	needs retraining	A Developer claim: the paper says the model is built for adaptation, "enabling UniVLA to adapt efficiently to diverse tasks and embodiments with minimal downstream data"; the model card sends fine-tuning to the repository ("For finetuning on simulation benchmarks or your customized dataset").
Works across robots	yes	C Developer claim, abstract: "the generalist policy can be deployed to various robots through efficient latent action decoding". 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

OpenDriveLab describes UniVLA as a framework for learning cross-embodiment vision-language-action policies.
The UniVLA paper by OpenDriveLab was submitted to arXiv on 9 May 2025.
OpenDriveLab titles its website Embodied AI and Autonomous Driving at HKU.

Sources

A arxiv.org/html/2505.06111v3 · Retrieved 2026-10-15
B huggingface.co/qwbu/univla-7b · Retrieved 2026-10-15
C arxiv.org/abs/2505.06111 · 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.