IntervenGen (I-Gen) is a data generation system that autonomously produces corrective interventions to improve robot policy robustness and data efficiency in imitation learning. The system uses a small number of human interventions to generate a large set of interventions covering diverse scene configurations and policy mistake distributions. I-Gen was tested on four simulated environments and one physical environment with object pose estimation error, achieving up to 39× improvement in policy robustness with only 10 human interventions. The system generates synthetic data by replaying human interventions and adapting them to new environments, allowing for efficient data collection and policy training. I-Gen outperforms alternative data collection methods, such as using full human demonstrations or MimicGen, by generating more diverse and effective interventions. The system also facilitates sim-to-real transfer of learned policies, showing that policies trained on synthetic data retain robustness to erroneous state estimation in real-world deployment. I-Gen is particularly effective in tasks involving object pose estimation errors, such as Nut Insertion, 2-Piece Assembly, and Coffee. The system's ability to generate diverse interventions reduces the need for extensive human supervision and improves policy performance across various environments and sources of observation error. I-Gen is a valuable tool for improving robot policies in high-precision, contact-rich tasks and for enabling efficient sim-to-real transfer of imitation learning policies.IntervenGen (I-Gen) is a data generation system that autonomously produces corrective interventions to improve robot policy robustness and data efficiency in imitation learning. The system uses a small number of human interventions to generate a large set of interventions covering diverse scene configurations and policy mistake distributions. I-Gen was tested on four simulated environments and one physical environment with object pose estimation error, achieving up to 39× improvement in policy robustness with only 10 human interventions. The system generates synthetic data by replaying human interventions and adapting them to new environments, allowing for efficient data collection and policy training. I-Gen outperforms alternative data collection methods, such as using full human demonstrations or MimicGen, by generating more diverse and effective interventions. The system also facilitates sim-to-real transfer of learned policies, showing that policies trained on synthetic data retain robustness to erroneous state estimation in real-world deployment. I-Gen is particularly effective in tasks involving object pose estimation errors, such as Nut Insertion, 2-Piece Assembly, and Coffee. The system's ability to generate diverse interventions reduces the need for extensive human supervision and improves policy performance across various environments and sources of observation error. I-Gen is a valuable tool for improving robot policies in high-precision, contact-rich tasks and for enabling efficient sim-to-real transfer of imitation learning policies.