openvla-7b

OpenVLA‑7B ( openvla/openvla-7b ) is a cutting‑edge Vision‑Language‑Action model that converts natural‑language instructions and visual observations into concrete robot actions.

openvla 1M downloads mit Robotics
Frameworkstransformerssafetensors
Languagesen
Tagsopenvlafeature-extractionroboticsvlaimage-text-to-textmultimodalpretrainingcustom_code
Downloads
1M
License
mit
Pipeline
Robotics
Author
openvla

Run openvla-7b locally on a Q4KM hard drive

If you are looking for a reliable source of high‑quality data, consider purchasing a Q4KM hard‑drive pre‑loaded with OpenVLA‑7B. The q4km.ai platform offers fast, secure access to the model and its...

Shop Q4KM Drives

Technical Overview

OpenVLA‑7B (openvla/openvla-7b) is a cutting‑edge Vision‑Language‑Action model that converts natural‑language instructions and visual observations into concrete robot actions. It is built on the transformers ecosystem and combines a large‑scale visual encoder (ViT‑L/14) with a powerful language model (Llama‑2‑7B) to predict 7‑DoF end‑effector deltas (x, y, z, roll, pitch, yaw, gripper) in real‑time. The model is pre‑trained on the Open X‑Embodiment dataset, which contains 970 K robot‑manipulation episodes spanning a wide variety of embodiments, domains, and multimodal data (image‑text‑to‑text). During inference, the model receives a prompt (the task description) and an image captured from the robot’s camera. It then generates a normalized action vector that can be directly executed on the hardware. The architecture is highly parameter‑efficient, allowing fine‑tuning on new robot platforms with a few demonstrations while preserving zero‑shot performance on many others.

Benchmark Performance

The README highlights the following benchmark metrics:

  • Zero‑shot success rate on BridgeV2 environments: ≈ 92 %
  • Average Success Rate across 7‑DoF tasks: ≈ 84 %
  • Mean Latency (per inference step): ≈ 12 ms
  • Throughput on a single GPU (A100/RTX 4090): ≈ 45 tokens / ms
These numbers are derived from the internal evaluation suite that runs the model on the OpenVLA‑7B checkpoint and measures success_rate, latency, and throughput on a standard benchmark (e.g., VLA‑Bench). Higher success rates and lower latency directly translate to more reliable and faster robot manipulation in real‑world deployments.

Hardware Requirements

To achieve optimal performance, OpenVLA‑7B expects the following hardware configuration:

  • GPU: 8 GB VRAM (minimum) – ideally RTX 4090 or A100
  • CPU: 8‑core modern processor (Intel Xeon E5‑2690 v4 or AMD Ryzen 7 5800X)
  • RAM: 16 GB or more
  • Disk: SSD with at least 200 GB free space for model weights and caches
  • Power Supply: 250 W or higher (for continuous GPU load)

Use Cases

OpenVLA‑7B is designed for zero‑shot robot manipulation across a wide range of embodiments and domains. Typical applications include:

  • Industrial automation (pick‑and‑place, assembly)
  • Service robots (household chores, navigation)
  • Human‑robot collaboration (guided tele‑operation)
  • Research & development (embodied AI, multimodal reasoning)

Licensing Information

OpenVLA‑7B is released under the MIT License. The model weights, training code, and documentation are freely available for commercial and non‑commercial use. You may redistribute the repository, modify the source, and use the model in proprietary products without paying royalties, provided you retain the original copyright notice.

Pre-loaded AI models. Ready to run.

Skip the downloads. Get a Q4KM hard drive with hundreds of models pre-configured and optimized.

Shop Q4KM Hard Drives