Real-to-sim-to-real data generation for manipulationIntermediate
AutoRSR: Automated Real-to-Sim-to-Real Data Generation for Scalable Robot Learning
Teaching a robot arm a task usually means a person guiding it through 100 or more demonstrations. This preprint scans the real workspace with a phone and two cameras, rebuilds it in simulation and generates 1,000 training runs automatically; a pi0.5 policy trained only on that synthetic data matched the success of one trained on 150 human demonstrations in two of three tasks.