GenSim leverages large language models to generate a vast number of robot simulation tasks, supporting both goal-oriented and exploratory generation modes. It can be used for multi-task strategy training and task-level generalization. By extending the existing benchmark by over 10 times using GPT-4, GenSim supports over 100 tasks. It generates code for robot simulation tasks through supervised fine-tuning and evaluation of multiple LLMs, including fine-tuned GPT and Code Llama. The multi-task strategies pre-trained on GPT-4 generated simulation tasks demonstrate significantly enhanced transferability in the real world, surpassing the baseline by over 25% after minimal simulation-to-real adaptation.