EmbodiedSWE Turns Frontier Coding Agents Into Generators of Diverse VLA Training Data

EmbodiedSWE, released Sept 24, uses frontier coding agents to attempt long-horizon robotics tasks including contact-rich manipulation, deformable-object handling, and operations requiring up to 30 minutes of continuous interaction. Agents solve individual instances but overfit; the paper's EmbodiedSWE-Gen expands a single agent solution into diverse trajectories used to train VLA policies, with sim-to-real transfer validated on physical robots and generalization scaling with demonstration count.