EmbodiedMidtrain: Bridging the Gap between Vision-Language Models and Vision-Language-Action Models via Mid-training
Updated
Updated · arxiv.org · Oct 16
EmbodiedMidtrain: Bridging the Gap between Vision-Language Models and Vision-Language-Action Models via Mid-training
1 articles · Updated · arxiv.org · Oct 16
Summary
Researchers have introduced EmbodiedMidtrain, a new mid-training framework to align Vision-Language Models (VLMs) with Vision-Language-Action Models (VLAs) for robotics.
The approach uses a proximity estimator to select VLM data most relevant to the embodied domain, significantly improving VLA performance on both simulated and real-world tasks.
This method bridges a key distribution gap, offering a stronger starting point for robot learning and is set to advance generalist robotic control capabilities.