Studies whether coding agents can generate scalable supervision for training general robot policies by introducing EmbodiedSWE-Bench, a simulation benchmark covering contact-rich manipulation and long-horizon tasks requiring up to 30 minutes of continuous interaction. The authors find frontier coding agents can solve these tasks and transfer solutions across tasks and embodiments, then introduce EmbodiedSWE-Gen to expand a single coding-agent solution into large, diverse trajectory sets for training vision-language-action (VLA) models. A VLA fine-tuned solely on coding-agent-generated simulation demonstrations successfully completes a long-horizon task on a real robot.
