MMaDA-VLA: Large Diffusion Vision-Language-Action Model with Unified Multi-Modal Instruction and Generation

TL;DR

MMaDA-VLA introduces a unified diffusion-based model for vision-language-action tasks, enabling consistent long-horizon robot control from visual and language inputs without extra modules.

Contribution

It proposes a native discrete diffusion framework that jointly models multi-modal understanding and generation in a single, unified architecture for robot manipulation.

Findings

Achieves 98.0% success on LIBERO benchmark.

Demonstrates state-of-the-art performance in real-world tasks.

Improves long-horizon consistency through iterative denoising.

Abstract

Vision-Language-Action (VLA) models aim to control robots for manipulation from visual observations and natural-language instructions. However, existing hierarchical and autoregressive paradigms often introduce architectural overhead, suffer from temporal inconsistency and long-horizon error accumulation, and lack a mechanism to capture environment dynamics without extra modules. To this end, we present MMaDA-VLA, a fully native pre-trained large diffusion VLA model that unifies multi-modal understanding and generation in a single framework. Our key idea is a native discrete diffusion formulation that embeds language, images, and continuous robot controls into one discrete token space and trains a single backbone with masked token denoising to jointly generate a future goal observation and an action chunk in parallel. Iterative denoising enables global, order-free refinement, improving long-horizon consistency while grounding actions in predicted future visual outcomes without auxiliary world models. Experiments across simulation benchmarks and real-world tasks show state-of-the-art performance, achieving 98.0% average success on LIBERO and 4.78 average length on CALVIN.