Open Source RoboticsRobots, software & learning

OpenVLA

Robotics model

Vision-language-action model code and checkpoints for image/instruction-conditioned robot action prediction. [1] [2] [8] [4] [5] [6] [7]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Robotics developers evaluating the selected source workflow [1] [2] [8] [4] [5] [6] [7]
Parameters
7,541,237,184 (~7.54B BF16 per HF safetensors metadata) [9]
Embodiments
WidowX (BridgeData V2), Open X-Embodiment pretraining robots, LIBERO simulation (Franka, via LoRA fine-tuning) [1] [7]
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

OpenVLA (openvla/openvla-7b) is a 7,541,237,184-parameter (~7.54B) Vision-Language-Action model pretrained on 970K robot manipulation episodes from the Open X-Embodiment dataset. Built on the Prismatic prism-dinosiglip-224px backbone—fusing DINOv2 ViT-L/14 and SigLIP ViT-So400M/14 vision encoders with a Llama-2 7B language model—it takes a single RGB camera image and an English instruction and autoregressively predicts discrete tokenized robot actions. [1] [7] [9]

Model card summary

Architecture

Prismatic VLM (fused DINOv2 ViT-L/14 + SigLIP ViT-So400M/14 vision backbone paired with Llama-2 7B language model) predicting discrete tokenized seven-component robot actions autoregressively.

[7] [9]

Base model

prism-dinosiglip-224px (DINOv2 ViT-L/14 + SigLIP ViT-So400M/14 + Llama-2 7B)

[7]

Parameter count

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

Verified via Hugging Face API safetensors metadata for openvla/openvla-7b.

[9]

Input modalities
rgblanguage

Single workspace RGB camera image and English task instruction.

[7]

Output modalities
ee-pose-deltagripper-command

Seven-component action: six end-effector pose deltas (x, y, z, roll, pitch, yaw) and one gripper command.

[7]

Action space

Single-step seven-component discrete tokenized action comprising six end-effector pose deltas (x, y, z, roll, pitch, yaw) and one gripper command, unnormalized per robot/dataset via unnorm_key (e.g., bridge_orig).

[1] [7]

Trained / supported embodiments
  • WidowX (BridgeData V2)
  • Open X-Embodiment pretraining robots
  • LIBERO simulation (Franka, via LoRA fine-tuning)

[1] [7]

Training datasets
  • Open X-Embodiment (970K robot manipulation episodes; component list in openvla/openvla)

[1] [7]

Interfaces
  • Hugging Face Transformers (AutoProcessor / AutoModelForVision2Seq with trust_remote_code=True)
  • HTTP deployment script (vla-scripts/deploy.py)

[7] [8]

Checkpoints

CheckpointFormatSizeAccessPinned revision
openvla/openvla-7bsafetensors14.05 GBopen47a0ec7fc4ec
openvla/openvla-7b-prismaticpytorch28.09 GBopen5e44aaf23f99

Ungated checkpoints verified via Hugging Face API; openvla-7b is for HF inference/LoRA and openvla-7b-prismatic is for full fine-tuning.

[9] [10]

Is this for you?

Use cases
  • Zero-shot control on embodiments represented in Open X-Embodiment (e.g., BridgeData V2 WidowX)
  • Parameter-efficient LoRA fine-tuning or full Prismatic fine-tuning on custom robot demonstrations

[1] [7]

Known limitations
  • Code MIT and model-card MIT declarations coexist with README warning about inherited Llama 2 model terms; no blanket weight clearance.
  • Bridge action normalization and robot-domain fine-tuning matter; predicted action is not a device command or proof of task success.
  • No weights downloaded/loaded, inference, fine-tuning, benchmark or robot actuation tested.

Editorial assessment of bounded source evidence; no runtime, fit, manufacturing or compatibility proof.

[1] [2] [8] [4] [5] [6] [7] · Editorial assessment

Safety and limitations
  • Code and model card declare MIT, while the repository README warns that pretrained weights inherit underlying Llama 2 Community License terms; no blanket weight clearance.
  • Zero-shot control is limited to represented embodiment/domain combinations; unseen embodiments or setups require collecting demonstrations and fine-tuning, plus per-robot, per-dataset action unnormalization statistics (unnorm_key).
  • HF inference requires trust_remote_code=True and a separate robot controller to interpret and execute predicted seven-component actions.

[1] [2] [7]

Compute & fine-tuning

Inference compute
  • Documented hardware: NVIDIA A100 (upstream LIBERO evals; CUDA bfloat16 quickstart)
  • Precision: bf16
  • Edge / consumer capable: No

14.05 GB BF16 safetensors footprint; exact minimum inference VRAM and latency_ms not measured here.

[1] [7]

Fine-tuning
  • Supported: Yes
  • Methods: LoRA (vla-scripts/finetune.py; ~27-72 GB on A100 80GB); Full fine-tuning (vla-scripts/train.py on 8x A100 with openvla-7b-prismatic)
  • Minimum VRAM: 27 GB
  • Recipe: Upstream fine-tuning documentation

Full fine-tuned Prismatic checkpoints require conversion via convert_openvla_weights_to_hf.py for HF inference.

[1]

Host computer

CUDA GPU (bfloat16 inference example; upstream LoRA fine-tuning example uses an A100 80 GB with ~72 GB at batch 16 or ~27 GB with smaller batches).

[1] [7]

Operating system
  • Linux

Standard CUDA/PyTorch Linux workflow documented.

[1] · Editorial assessment

Dependencies
  • Python 3.10 reference environment; package >=3.8 with pinned torch 2.2/transformers 4.40.1 stack
  • GPU/precision and attention dependencies selected for inference versus fine-tuning; robot controller/data separate

[1] [2] [8] [4] [5] [6] [7]

Skills
  • Python
  • PyTorch
  • Hugging Face Transformers
  • RLDS / robot action unnormalization

Inferred from documented Python/PyTorch/Transformers/RLDS setup.

[1] [7] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [1] [2] [8] [4] [5] [6] [7]

  • Inspect the pinned implementation and select the exact configuration described in the article.
  • Follow the chosen upstream environment/build route, resolving the listed dependencies before runtime or hardware integration.

Editorial assessment of bounded source evidence; no runtime, fit, manufacturing or compatibility proof.

[1] [2] [8] [4] [5] [6] [7] · Editorial assessment

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 (Spatial, Object, Goal, Long)76.5 ± 0.6% average (Spatial 84.7 ± 0.9%, Object 88.4 ± 0.8%, Goal 79.2 ± 1.0%, Long 53.7 ± 1.3%)Upstream-reported LoRA (r=32) fine-tuning results averaged over 3 random seeds x 500 rollouts on modified LIBERO datasets; not independently verified by OSR.

[1]

Project notes

Action contract and checkpoint variants

The model predicts a single-step seven-component action comprising six end-effector pose deltas (x, y, z, roll, pitch, yaw) and one gripper command, which must be unnormalized using per-robot, per-dataset statistics (unnorm_key, such as bridge_orig for BridgeData V2 WidowX). Two ungated checkpoints are published on Hugging Face Hub: openvla/openvla-7b (14.05 GB BF16 safetensors) for Transformers inference (trust_remote_code=True) and LoRA fine-tuning, and openvla/openvla-7b-prismatic (28.09 GB PyTorch .pt) for full fine-tuning via vla-scripts/train.py, whose outputs must be converted back to Hugging Face format using convert_openvla_weights_to_hf.py. Camera capture and robot.act calls in the quickstart are placeholders; a separate robot controller is required. [1] [7] [8] [9] [10]

Fine-tuning compute and upstream-reported benchmarks

The reference environment uses Python 3.10 (package metadata allows >=3.8) with pinned torch 2.2.0 and transformers 4.40.1. Documented zero-shot use is limited to represented embodiment/domain combinations from the pretraining mixture; upstream explicitly directs users with unseen robot embodiments or setups to collect target demonstrations and fine-tune. As a concrete custom-data step, prepare demonstrations in RLDS format and register the dataset configuration and transform, or inspect vla-scripts/finetune.py's custom PyTorch Dataset wrapper for LoRA. The documented inference example runs on CUDA in bfloat16, but minimum inference VRAM and latency are not established here—checkpoint download sizes (14.05–28.09 GB) are disk sizes rather than runtime memory requirements. For training, the LoRA example runs on an NVIDIA A100 80 GB (~72 GB at batch size 16, or ~27 GB with smaller batches), while full fine-tuning targets an 8-GPU A100 node. In the repository README, upstream reports a 76.5 ± 0.6% average success rate on modified LIBERO suites after LoRA (r=32) fine-tuning (Spatial 84.7 ± 0.9%, Object 88.4 ± 0.8%, Goal 79.2 ± 1.0%, Long 53.7 ± 1.3%, averaged over 3 seeds x 500 rollouts). All benchmark scores and VRAM numbers are upstream-reported; Open Source Robotics has not downloaded weights or run training, fine-tuning, inference, or robot evaluations. [1] [5] [6] [7]

What we checked

Pinned source and selected artifact structure review only; no dependency installation, upstream build, native CAD/kernel/fit/strength validation, fabrication, wiring, calibration, simulation, powered operation or performance testing. Original archives stay private.

Not yet established

  • Documented closed-loop control frequency (Hz)

Updates

Upstream last updated
Feb 17, 2026 [9]
Latest release
Not established
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
c8f03f4 · combined hardware and software repository

Separate license matrix

Distinct inspectable robotics software/data workflow with specific contribution and bounded next source-investigation step; not performance or physical-build certification.

LayerAvailability / license
Codeavailable · MIT
Selected project implementation; external dependencies, hardware, weights/data and asset rights remain separate. [8] [2]
Model weightsavailable · MIT declaration; inherited Llama-2 terms noted separately
Officialcard declaresMIT; codeREADME warns underlyingLlama2terms; modelcard/metadata inspected only, no weight download/load or blanketrights conclusion. [7]
Training datasetsavailable · Mixed (Open X-Embodiment constituent dataset licenses apply)
Pretrained on 970K episodes from Open X-Embodiment; individual constituent dataset licenses apply. [1] [7]
Documentationavailable · MIT
Scoped project declaration; no blanket grant over all linked material. [1] [2]

Layered code, weight, and dataset rights

The openvla/openvla repository is MIT-licensed, and the pinned openvla/openvla-7b model card also declares MIT for checkpoints. However, the code README separately notes that pretrained models inherit underlying Llama 2 Community License terms from the Llama-2 backbone. Both declarations are preserved here without offering blanket commercial clearance. Constituent Open X-Embodiment training datasets retain their own upstream licenses. [1] [2] [7]

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 — Migrated 0.7.0-draft.5 -> 0.8.0-draft.1 (kind=model); added verified HF API metadata for openvla-7b (S009) and openvla-7b-prismatic (S010) while preserving historical source dates.