Physics-aligned simulation can replace expensive real-world data collection for deformable object manipulation—synthetic data from calibrated digital twins achieves 90% zero-shot success and matches real-data baselines at a 1:15 data efficiency ratio.
SIM1 creates physics-accurate digital twins of deformable objects from real demonstrations, then generates synthetic training data through simulation to train robotic manipulation policies. The system achieves real-world performance comparable to policies trained on 15x more real data, showing that properly grounded simulation can solve the data scarcity problem in soft object manipulation.