Synthetic Data Generation for Robotics

This field studies how to create useful robot-training and evaluation data without collecting every example on physical hardware, supporting tasks from manipulation and navigation to humanoid motion and tactile sensing. Common approaches include physics-based simulators with procedural scenes and domain randomization, automated trajectory planning, retargeting human demonstrations, and generative video or world models that produce observations paired with actions. Researchers test these methods in settings such as tabletop manipulation, deformable-object handling, and navigation, often measuring transfer to real robots as well as performance on simulation benchmarks. The central tension is scaling data and visual diversity while keeping scenes, sensor signals, and actions physically consistent; realistic-looking outputs alone do not guarantee useful training data or reliable sim-to-real transfer.