Text-to-Robot Simulation and Policy Training

eigenron · x · 2026-07-16

The post demonstrates a text-to-robot training pipeline: using Gizmo to generate a pick-and-place scene in about 10 minutes, exporting it as MJCF, and integrating it with the YAM arm.

The author claims to have subsequently generated 500 episodes and trained an ACT policy within a few hours. They also noted that this workflow can "one-click" convert real images/videos into simulation scenes for MuJoCo / Isaac, and invited users to reach out to try it.

Related event: Gizmo: Generate Robot Simulation Scenes from Text in Minutes(4 posts)→

Original post →

More from Embodied

Embodied channel →