EmbodiedSmith: Scaling Embodied Data through Recursive Self-Improvement Flywheel in Simulation
EmbodiedSmith uses a recursive simulation loop to make robot training data broader and more useful.
The framework links asset, scene, and task generation so each can refine the others. Its scene generator anticipates task needs, while task generation points back to scene edits, improving success on generated tasks, including long-horizon ones. The paper says it supports mobile manipulators, humanoids, dexterous hands, deformable objects, and fluids. Experiments report better data quality, diversity, efficiency, and stronger downstream generalization from more diverse data. HF Daily Papers' note
The framework links asset, scene, and task generation so each can refine the others. Its scene generator anticipates task needs, while task generation points back to scene edits, improving success on generated tasks, including long-horizon ones. The paper says it supports mobile manipulators, humanoids, dexterous hands, deformable objects, and fluids. Experiments report better data quality, diversity, efficiency, and stronger downstream generalization from more diverse data. HF Daily Papers' note
score 5