GEM: Generative Supervision Helps Embodied Intelligence

Researchers introduced GEM, a vision-language model that adds depth map generation during pretraining to better learn spatial and physical cues for robotics.

They also released GEM-4M, a dataset with grounding, reasoning, planning, and depth supervision data to support embodied learning.

GEM and GEM-VLA achieved state-of-the-art results on embodied benchmarks and showed strong performance in both real-world and simulated robot tasks.

The work highlights how adding spatial supervision to vision-language pretraining can close a key gap in robotics.