GEM: Generative Supervision Helps Embodied Intelligence

Researchers introduced GEM, a generative-supervised embodied vision-language model that adds depth-map generation to pretraining.

The team also released GEM-4M, a dataset with grounding, reasoning, planning, and depth supervision for embodied learning.

GEM achieved state-of-the-art results on multiple embodied benchmarks.

A GEM-based vision-language-action model, GEM-VLA, showed better simulation and real-world task execution, suggesting depth supervision improves spatial understanding and robot control.