Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Yang Liu, Pengxiang Ding, Tengyue Jiang, Xudong Wang, Wenxuan Song, Minghui Lin, Han Zhao, Hongyin Zhang, Zifeng Zhuang, Wei Zhao, Siteng Huang, Jinkui Shi, Donglin Wang

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.

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

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.

Paper Structure

This paper contains 30 sections, 5 equations, 6 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Large Diffusion Models
  4. Vision-Language-Action Models
  5. World Model for Robotics
  6. Methodology
  7. Problem Formulation
  8. Architecture
  9. Data Tokenization.
  10. Multi-Modal Sequence Modeling.
  11. Hybrid Attention Mechanism.
  12. Learning Objective
  13. Inference
  14. Iterative Denoising.
  15. Key-Value Cache.
  16. ...and 15 more sections

Figures (6)

  • Figure 1: (a) Conventional VLA architectures vs. MMaDA-VLA. (b) Comparison with discrete diffusion methods. (c) Performance comparison with SOTAs. Cur., Obs., Gen., and DDVLA denote current, observation, generation, and Discrete Diffusion VLA.
  • Figure 2: Schematic of MMaDA-VLA framework and training pipeline.
  • Figure 3: Real-world experiment setup, task examples and performance comparison.
  • Figure 4: MMaDA-VLA execution examples on real-world tasks.
  • Figure 5: Pre-training visualization.
  • ...and 1 more figures