Dexbotic (Dexmal VLA toolbox)

An open-source VLA training-and-deployment toolbox from Dexmal that brings several mainstream VLA methods into one codebase.

Dexbotic is an open-source, PyTorch-based toolbox for VLA (vision-language-action model) development, released by Dexmal (原力灵机). Different VLA papers each tend to come with their own repository, data format, and training scripts, which makes reproducing and comparing them a lot of work; Dexbotic folds several mainstream VLA methods into a single unified framework, with a common data interface, pretrained weights, and shared training and evaluation pipelines, so researchers can swap models more easily for experiments, or fine-tune on their own robot data and deploy. It sits in the same category as LeRobot, openpi, and starVLA — a “VLA code base” — and exactly which models it supports is best checked against the official repository.

ExampleConvert your own dual-arm data into Dexbotic's data format, fine-tune one of its built-in VLA configurations, then deploy to a real robot using the inference script the repository provides.

Related

Dexmal · Vision-Language-Action Model · LeRobot · openpi (Physical Intelligence) · starVLA · Fine-tuning

Sources

Dexbotic GitHub

As of

2026-09