Open Source RoboticsRobots, software & learning

UniVLA

Robotics model

7.54B-parameter generalist Vision-Language-Action model (RSS 2025) pretrained on Open X-Embodiment and Ego4D human videos via task-centric embodiment-agnostic latent actions, with Apache-2.0 code and weights. [1] [2] [3]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Robotics researchers; Cross-embodiment VLA developers; Video-to-action pretraining teams [4]
Parameters
7,541,237,184 (~7.54B BF16 parameters per safetensors metadata) [2]
Embodiments
Franka Panda (LIBERO / CALVIN), WidowX 250 (BridgeV2 / SimplerEnv), Google Robot (SimplerEnv), … [4]
Source rights
Named licenses are documented by part. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware

Requirements · Research details and dates

On this page

Overview

UniVLA (qwbu/univla-7b) is a 7,541,237,184-parameter (~7.54B) generalist Vision-Language-Action model (RSS 2025) trained in two stages: a task-centric Latent Action Model (LAM) paired with a 7B Prismatic VLM backbone (TRI-ML/prismatic-vlms/prism-dinosiglip-224px+7b, combining DINOv2 and SigLIP vision encoders with a 7B LLM). Pretraining draws on Open X-Embodiment robot demonstrations and Ego4D human egocentric videos. [1] [2] [4]

Model card summary

Architecture

Two-stage task-centric Latent Action Model (LAM) paired with a 7B Prismatic VLM backbone (TRI-ML/prismatic-vlms prism-dinosiglip-224px+7b, OpenVLA-style DINOv2 + SigLIP vision encoders + 7B LLM) predicting discrete latent action tokens decoded into robot actions during post-training.

[2] [4]

Base model

TRI-ML/prismatic-vlms/prism-dinosiglip-224px+7b

Explicitly listed in UniVLA README Model Zoo table.

[4]

Parameter count

7,541,237,184 (~7.54B BF16 parameters per safetensors metadata)

Exact tensor count verified via Hugging Face API safetensors metadata.

[2]

Input modalities
rgblanguage

RGB visual observations (224px DINOv2+SigLIP) and language task instructions.

[2] [4]

Output modalities
ee-pose-deltagripper-commandwaypoints

Pretrained model outputs embodiment-agnostic latent action tokens; post-trained policies decode to manipulation actions (LIBERO/CALVIN/SimplerEnv/real-world) or navigation actions (Room2Room).

[4]

Action space

Unified discrete task-centric latent action tokens during pretraining, decoded to embodiment-specific continuous manipulation or navigation action spaces via post-training action decoder.

[2] [4]

Trained / supported embodiments
  • Franka Panda (LIBERO / CALVIN)
  • WidowX 250 (BridgeV2 / SimplerEnv)
  • Google Robot (SimplerEnv)
  • Navigation agent (Room2Room)
  • Real-world tabletop arm (per README demo videos)

[4]

Training datasets

Full univla-7b pretrained on OpenX + Ego4D (manipulation, navigation, and human videos).

[2] [4]

Interfaces
  • Hugging Face Transformers safetensors
  • Python training & evaluation scripts

[2] [4]

Checkpoints

CheckpointFormatSizeAccessPinned revision
qwbu/univla-7b (full OpenX + Ego4D pretrained)safetensors14.05 GBopen968e2896809c

Additional checkpoints (univla-latent-action-model, univla-7b-bridge-pt, univla-7b-human-pt, univla-libero) are linked in the README Model Zoo.

[2] [4]

Is this for you?

Use cases
  • Pretraining VLAs from unlabeled human and cross-embodiment videos
  • Post-training/fine-tuning on manipulation and navigation benchmarks (LIBERO, CALVIN, SimplerEnv, Room2Room)
  • Real-robot manipulation transfer

[4]

Known limitations
  • Pretrained univla-7b plans in latent action space and requires post-training / action-decoder adaptation for a specific robot embodiment
  • Backbone derives from Prismatic VLM (prism-dinosiglip-224px+7b / Llama-2 family), so inherited Llama-2 terms must be reviewed for commercial deployment

[2] [4]

Safety and limitations
  • Base checkpoint operates in latent action space and must be post-trained with an embodiment-specific action decoder before hardware control.
  • Upstream terms of the Prismatic VLM backbone and pretraining datasets (Open X-Embodiment, Ego4D) were not excerpted in this review; impact of Ego4D dataset-access terms on checkpoint-only deployment is not established here.

Inferred from upstream model card scope, pretraining distribution, and documented deployment prerequisites.

[2] [4] · Editorial assessment

Compute & fine-tuning

Inference compute
  • Precision: bf16
  • Edge / consumer capable: No

Exact min_vram_gb and latency_ms not stated on model card; 14.05 GB BF16 weights preclude small edge devices without quantization.

[2] · Editorial assessment

Fine-tuning
  • Supported: Yes
  • Methods: embodiment-specific action decoder (~12M params) + LoRA (rank 32, ~123M total trainable params in upstream recipe)
  • Recipe: Upstream fine-tuning documentation

[4]

Host computer

14.05 GB BF16 checkpoint requires >=16–24 GB VRAM GPU for inference; exact minimum VRAM and multi-GPU fine-tuning specs not pinned in HF card.

Inferred from 7.54B BF16 safetensors footprint (14.05 GB).

[2] · Editorial assessment

Operating system
  • Linux

Inferred from documented Linux/CUDA Python environment and shell instructions.

[4] · Editorial assessment

Dependencies
  • Python
  • PyTorch
  • Transformers
  • Prismatic VLM / OpenVLA architecture dependencies

[2] [4]

Skills
  • Python
  • PyTorch
  • Transformers / OpenVLA codebase familiarity
  • Multi-GPU fine-tuning

Inferred from documented Python/PyTorch/CLI setup and training dependencies.

[2] [4] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [4] · Model card / documentation [1] [2]

  • Clone OpenDriveLab/UniVLA and follow installation instructions
  • Download qwbu/univla-7b (and optionally qwbu/univla-latent-action-model) from Hugging Face
  • Post-train on LIBERO, CALVIN, Room2Room, SimplerEnv, or custom robot dataset

[2] [4]

Upstream-reported benchmarks

Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.

BenchmarkUpstream-reported scoreScope & caveats
LIBERO, CALVIN, Room2Room, SimplerEnv, real-world manipulation (arXiv:2505.06111)See OpenDriveLab/UniVLA README & arXiv:2505.06111Upstream-reported simulation and real-robot evaluation claims; not independently verified by OSR.

[4]

Project notes

Latent action space requires post-training adaptation

During pretraining, UniVLA predicts discrete embodiment-agnostic latent action tokens rather than direct robot joint commands. Deploying the ungated BF16 safetensors checkpoint on a physical arm or navigation platform follows a staged workflow documented in the pinned README's post-training section: start from the pretrained latent-action policy, attach an embodiment-specific action decoder head, and run downstream fine-tuning on target demonstrations. In upstream's reported recipe, the action decoder adds roughly 12M parameters and is trained alongside LoRA (rank 32) for approximately 123M total trainable parameters; treat these as upstream recipe figures rather than universal requirements, and do not infer minimum VRAM or training cost from them. Upstream does not pin a single minimum VRAM or closed-loop latency figure in the model card, so hardware sizing must be validated locally before closed-loop execution. [2] [4]

Upstream-reported benchmark coverage

The repository README and arXiv:2505.06111 report evaluations across LIBERO, CALVIN, Room2Room navigation, SimplerEnv, and real-world tabletop manipulation tasks. All reported results are strictly upstream-reported; Open Source Robotics has not executed training, post-training, inference, or robot trials for UniVLA and does not independently verify task performance or control rates. [4]

Separate code, weights, backbone, and dataset terms

The OpenDriveLab/UniVLA GitHub repository carries a verified root Apache-2.0 LICENSE, and the qwbu/univla-7b Hugging Face model card declares apache-2.0. Separate from those confirmed Apache-2.0 declarations, the pretrained checkpoint derives from TRI-ML/prismatic-vlms/prism-dinosiglip-224px+7b and pretraining draws on Open X-Embodiment and Ego4D. The upstream terms of that backbone and those datasets were not excerpted in this review and require independent verification; in particular, this review does not establish how Ego4D's dataset-access terms affect checkpoint-only deployment. [1] [2] [3] [4]

What we checked

None. Prepared from public upstream sources without hands-on hardware execution.

Not yet established

  • Documented closed-loop control frequency (Hz)

Updates

Upstream last updated
Nov 19, 2025 [1]
Latest release
Aug 2, 2025 [2]
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
May 21, 2027
Reviewed revision
OpenDriveLab/UniVLA@0ab9e9dd3074ffa14fdd1a00aa2aa84e4f8e1352; HF qwbu/univla-7b@968e2896809c00968e8fd936e052476ce5ac1bd0 · software repository

Separate license matrix

Open-weights 7.54B generalist VLA with Apache-2.0 codebase and ungated safetensors weights demonstrating cross-embodiment latent-action pretraining; inherited backbone and Ego4D data terms disclosed explicitly.

LayerAvailability / license
Codeavailable · Apache-2.0
Apache License Version 2.0 verified in root LICENSE. [1] [3]
Model weightsavailable · Apache-2.0 (HF card declaration; upstream terms of base backbone TRI-ML/prismatic-vlms/prism-dinosiglip-224px+7b require independent verification)
HF card declares license: apache-2.0; fine-tuned from TRI-ML/prismatic-vlms/prism-dinosiglip-224px+7b, whose upstream terms were not excerpted in this review. [2] [4]
Training datasetsavailable · Mixed upstream (Open X-Embodiment + Ego4D; upstream dataset terms not excerpted in this review)
[2] [4]
Documentationavailable · Apache-2.0
[3] [4]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history

2026-11-21 — Initial 0.8.0-draft.1 model draft prepared from pinned GitHub commit 0ab9e9dd3074 and HF revision 968e2896809c.