AutoSceneGen
A universal, configurable, and cost-effective framework used in autonomous driving research to automatically generate diverse, safety-critical traffic scenarios for simulation, testing and safety evaluation of autonomous vehicles. Proposed by computer scientist Aizierjiang Aiersilan at the AAAI-25 conference in the paper "Generating Traffic Scenarios via In-Context Learning to Learn Better Motion Planner," the framework leverages the in-context learning (ICL) capabilities of existing Large Language Models (LLMs) to create traffic scenarios for testing autonomous driving systems.
By using the AutoSceneGen framework, developers bypass the need to manually program simulators or train separate generative models, significantly reducing the human effort and budget required to create complex, rare, or accident-prone traffic situations—often called corner cases—which are essential for training robust autonomous motion planners, as demonstrated in the original paper.
By using the AutoSceneGen framework, developers bypass the need to manually program simulators or train separate generative models, significantly reducing the human effort and budget required to create complex, rare, or accident-prone traffic situations—often called corner cases—which are essential for training robust autonomous motion planners, as demonstrated in the original paper.
Engineer 1: "Our motion planner needs more training data on how to react to wrongly parked cars with open doors in the rain."
Engineer 2: "Don't write the code manually. Just write a quick text prompt and run it through AutoSceneGen so the LLM can generate the exact simulation scripts we need for the CARLA engine."
Engineer 2: "Don't write the code manually. Just write a quick text prompt and run it through AutoSceneGen so the LLM can generate the exact simulation scripts we need for the CARLA engine."
AutoSceneGen by sliverinc February 21, 2026
