Discrete Diffusion VLA: Bringing Discrete Diffusion to Action Decoding in Vision-Language-Action Policies

TL;DR

This paper addresses the bottleneck in Vision-Language-Action policies arising from autoregressive or decoupled action decoders by introducing Discrete Diffusion VLA, a unified transformer that performs discrete diffusion over fixed-length action chunks inside the VLM backbone. It enables adaptive easy-first decoding with secondary re-masking, retaining pretrained vision-language priors and achieving parallel refinement with fewer function evaluations. Empirically, it achieves state-of-the-art performance across LIBERO and SimplerEnv benchmarks and demonstrates robustness under OOD perturbations, while offering clear inference efficiency advantages over autoregressive approaches. The work paves a path toward scalable, unified multimodal foundation models that jointly handle vision, language, and action at scale.

Abstract

Vision-Language-Action (VLA) models adapt large vision-language backbones to map images and instructions into robot actions. However, prevailing VLAs either generate actions auto-regressively in a fixed left-to-right order or attach separate MLP or diffusion heads outside the backbone, leading to fragmented information pathways and specialized training requirements that hinder a unified, scalable architecture. We present Discrete Diffusion VLA, a unified-transformer policy that models discretized action chunks with discrete diffusion. The design retains diffusion's progressive refinement paradigm while remaining natively compatible with the discrete token interface of VLMs. Our method achieves an adaptive decoding order that resolves easy action elements before harder ones and uses secondary re-masking to revisit uncertain predictions across refinement rounds, which improves consistency and enables robust error correction. This unified decoder preserves pre-trained vision-language priors, supports parallel decoding, breaks the autoregressive bottleneck, and reduces the number of function evaluations. Discrete Diffusion VLA achieves 96.3% avg. success rates on LIBERO, 71.2% visual matching on SimplerEnv-Fractal and 54.2% overall on SimplerEnv-Bridge. We also provide ablation study on vision-language ability retention on LIBERO-OOD (Out-of-Distribution) benchmark, with our method improving over autoregressive, MLP decoder and continuous diffusion baselines. These findings indicate that discrete-diffusion VLA supports precise action modeling and consistent training, laying groundwork for scaling VLA to larger models and datasets. Our code is available at https://github.com/Liang-ZX/DiscreteDiffusionVLA/tree/libero.

Figures (15)

• Figure 1: Paradigm comparison. Continuous diffusion over action chunks (left) versus discrete token decoders: AR (sequential), BERT-style (parallel), and our discrete diffusion with re-masking.
• Figure 2: Discrete Diffusion VLA architecture. A single-transformer VLM backbone encodes multi-view RGB (SigLIP+DINOv2 ViTs) and a tokenized instruction, and decodes discrete action chunks via diffusion-style iterative refinement. Adaptive Decoding (bottom left) keeps high-confidence tokens each round and anneals the keep ratio with a cosine schedule for easy-first parallel refinement. Secondary Re-Masking (bottom right) uses threshold and residual-drop checks to re-mask uncertain tokens, enforcing cross-step consistency and robust error correction.
• Figure 3: Benchmarks and tasks. We evaluate Discrete Diffusion VLA across three robot settings: LIBERO with a Franka Panda arm, SimplerEnv–Fractal with a Google Robot, and SimplerEnv–Bridge with a WidowX Arm.
• Figure 4: Vision Augmentation Sample of LIBERO-OOD Goal Task. The out-of-distribution (OOD) setting introduces visual perturbations by replacing objects with variants that differ in scale, material properties, and visual appearance. Specifically, the modified scenes include a larger bowl and a stove with metallic luster reflections.
• Figure 5: Speed–Quality trade-off. (i) Time efficiency by the number of generated action chunks per second. (ii) Ablation on denoising steps. We adopt $T=12$ as a knee point for high accuracy at strong throughput.