DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Driving

TL;DR

This work identifies a supervision bottleneck in large Vision-Language-Action driving models and proposes DriveVLA-W0, which uses world-model-based future-image prediction as dense self-supervision. It supports both an autoregressive world model for discrete visual tokens and a diffusion-based world model for continuous features, paired with a lightweight Mixture-of-Experts action decoder for real-time inference. Across NAVSIM benchmarks and a massive 70M-frame in-house dataset, DriveVLA-W0 delivers state-of-the-art results and reveals that world modeling amplifies data scaling laws, enabling more effective transfer and generalization than action-only supervision. A key finding is that, at scale, simpler autoregressive action decoders outperform more complex flow-matching approaches, highlighting a surprising shift in decoding strategy when data grows massively.

Abstract

Scaling Vision-Language-Action (VLA) models on large-scale data offers a promising path to achieving a more generalized driving intelligence. However, VLA models are limited by a ``supervision deficit'': the vast model capacity is supervised by sparse, low-dimensional actions, leaving much of their representational power underutilized. To remedy this, we propose \textbf{DriveVLA-W0}, a training paradigm that employs world modeling to predict future images. This task generates a dense, self-supervised signal that compels the model to learn the underlying dynamics of the driving environment. We showcase the paradigm's versatility by instantiating it for two dominant VLA archetypes: an autoregressive world model for VLAs that use discrete visual tokens, and a diffusion world model for those operating on continuous visual features. Building on the rich representations learned from world modeling, we introduce a lightweight action expert to address the inference latency for real-time deployment. Extensive experiments on the NAVSIM v1/v2 benchmark and a 680x larger in-house dataset demonstrate that DriveVLA-W0 significantly outperforms BEV and VLA baselines. Crucially, it amplifies the data scaling law, showing that performance gains accelerate as the training dataset size increases.

Figures (14)

• Figure 1: World modeling as a catalyst for VLA data scalability. (a): Unlike standard VLAs trained solely on action supervision, our DriveVLA-W0 is trained to predict both future actions and visual scenes. (b): This world modeling task provides a dense source of supervision, enabling our model to better harness the benefits of large-scale data.
• Figure 2: The architecture of DriveVLA-W0, which achieves world modeling in two ways: (a) an AR World Model that predicts discrete visual tokens, and (b) a Diffusion World Model that denoises latent representations conditioned on multimodal inputs.
• Figure 3: (a) Our Mixture-of-Experts (MoE) architecture pairs a large VLA Expert with a lightweight Action Expert for efficient inference. (b-d) This framework serves as a testbed for comparing three action decoding schemes: query-based, autoregressive, and flow matching.
• Figure 4: World modeling unlocks generalization with data scaling. Our world model turns pretraining from a detriment for sparse-supervision baselines (red arrows) into a benefit for our VLA-W0s (green arrows), enabling positive knowledge transfer across datasets with similar visuals (b) but different action distributions (a). Figure (a) is from navsim.
• Figure 5: Action-Conditioned Simulation. We demonstrate the utility of our world model as a simulator capable of generating diverse futures based on control signals. In the recorded ground truth, the ego vehicle maintains speed and drives forward. By conditioning the model on a counterfactual "decelerate" action, the world model generates a distinct visual outcome where the vehicle slows down.