OpenVLA

Kim 2024 · Paper

One-line summary — OpenVLA is a 7B-parameter open-source vision-language-action model trained on 970k real-world robot demonstrations from Open X-Embodiment, matching or beating much larger closed VLAs while being efficiently fine-tunable on consumer GPUs.

Problem

Large policies pretrained on Internet-scale vision-language data plus diverse robot demonstrations promise a new way to teach robots: fine-tune a general vision-language-action (VLA) model instead of training each new behavior from scratch. But adoption was blocked on two fronts — existing VLAs (RT-2 and kin) were largely closed with limited visibility into architecture, training procedure, and data mixture, and prior work had not explored how to efficiently fine-tune VLAs for new tasks on commodity hardware, the key step for practical use. OpenVLA set out to remove both blockers.

Method & architecture

Results

Why it matters for SLAM

OpenVLA is the reference open VLA and the base model for a large body of follow-on robot-learning work, showing that a 7B open model can outperform far larger closed systems. For the SLAM community it matters in two ways: it validates SigLIP/DINOv2-style foundation features as robot perception backbones, and it defines the action side of the emerging autonomy stack — SLAM supplies localization and metric maps while a VLA closes the loop from perception and language to physical action.