Vision-language-action models (VLAs) inherit their visual and linguistic abilities from vision-language models (VLMs), but most are built from off-the-shelf VLMs that are not adapted to the embodied domain. The authors first characterize this gap, showing that VLA data occupy compact regions largely separated from the broader VLM distribution, with alignment varying substantially across and within VLM data sources.

To bridge the gap, they propose EmbodiedMidtrain, a mid-training stage before VLA fine-tuning. A lightweight learnable proximity estimator selects the most VLA-aligned candidates from a large VLM pool, producing the EmbodiedMidtrain-1M dataset. Mid-training the VLM on this data before downstream VLA fine-tuning yields consistent performance gains on simulated robot manipulation benchmarks and real-world tasks, across different VLM backbones and VLA action heads.

These gains also extend to downstream tasks not used to guide data selection. Analysis shows that mid-training provides a stronger initialization, with improvements emerging from the earliest training steps and widening throughout. 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 original VLM data. The authors plan to release all code, data, and models.