EmbodiedMidtrain: Bridging the Gap between Vision-Language Models and Vision-Language-Action Models via Mid-training

TL;DR

EmbodiedMidtrain enhances vision-language models for embodied tasks by selectively mid-training on VLA-aligned data, improving downstream robot manipulation performance.

Contribution

This work introduces a data-driven mid-training approach that bridges the gap between VLMs and VLAs, improving their suitability for embodied applications.

Findings

Mid-training improves VLA performance across multiple benchmarks.

The data engine effectively identifies VLA-aligned data, enhancing model initialization.

Mid-training benefits are evident from early training stages.

Abstract

Vision-Language-Action Models (VLAs) inherit their visual and linguistic capabilities from Vision-Language Models (VLMs), yet most VLAs are built from off-the-shelf VLMs that are not adapted to the embodied domain, limiting their downstream performance. In this work, we propose EmbodiedMidtrain to bridge the gap between VLMs and VLAs. We first characterize the data distribution gap between them, showing that VLA data occupy compact regions that are largely separated from the broader VLM distribution, while the degree of alignment varies substantially both across and within VLM data sources. Then, we build a mid-training data engine that leverages a lightweight learnable proximity estimator to select the most VLA-aligned candidates from a large VLM pool, and mid-trains the VLM on this curated mixture before downstream VLA fine-tuning. Experiments on three robot manipulation benchmarks show that mid-training consistently improves performance across different VLM backbones, achieving results competitive with expert VLAs and off-the-shelf VLMs trained with larger model scale and training budgets. Further analysis reveals that mid-training provides a stronger initialization for VLA fine-tuning, with gains emerging from the earliest steps and widening throughout training. Moreover, the data engine captures both dataset-level and sample-level alignment signals, favoring spatial reasoning over text-centric tasks while preserving the diversity of the VLM data. We will release all code, data and models for future research.