VLANeXt Offers an Open Recipe Book for Building Stronger Robot AI

Vision-language-action (VLA) systems connect camera input and natural-language instructions to robot movements. But comparing them has become difficult: research teams often change the visual-language backbone, policy design, action format, training method, and evaluation setup at the same time.

A Hugging Face Blog community article published August 31, 2026, introduces VLANeXt as an attempt to make those comparisons clearer. Rather than presenting one complicated architecture, the project provides an open codebase and a set of repeatable recipes for robotics foundation-model research. Its associated ICML work reports more than 500 experiments across the VLA design space.

What the experiments found

The team began with an RT-2/OpenVLA-style baseline and evaluated it using the Spatial suite in LIBERO and LIBERO-plus, including tests involving unseen perturbations. The resulting design changes include a dedicated policy module instead of reusing language tokens for actions, a stronger MetaQuery-based policy component, and prediction of eight future actions at once rather than only the next movement.

The project also compares different ways to represent actions, including discrete bins, VQ-VAE codebooks, regression, DDIM, and flow matching. That makes VLANeXt useful not only as a model implementation, but also as a controlled starting point for testing alternatives.

The expanded release covers multiple research directions and scales: VLANeXt-S, -L, and -XL; latent action learning through VLANeXt-LAM; latent-space predictive modeling with VLANeXt-JEPA; and world-action modeling through VLANeXt-WAM. Researchers can use these baselines to study how model size, action representation, and learning objective affect robot behavior without rebuilding every component from scratch.

For AI-tool builders, the significance is practical: VLANeXt packages robotics experimentation into a more accessible, modular workflow and may make future VLA results easier to reproduce and interpret.

Source: Hugging Face Blog