ADRIVE-GPT is dedicated to the research and application of anchored language models (Large Language Models, LLMs) in the context of automated driving. Previous studies have only considered LLMs for simple tasks in stationary robotics. ADRIVE-GPT will investigate how multimodal sensor perceptions from camera and lidar sensors can be combined with natural language to generate action plans for accident-free and human-like autonomous driving. In addition, it is being investigated whether the learning process can be supported by already processed knowledge about static and dynamic objects.

The project is also investigating whether traffic rules or detailed map information must be entered explicitly or whether the model can learn this information implicitly. The latter would enable the model to act more independently of fixed rules and information and make the model more scalable overall. Another focus of ADRIVE-GPT is on researching and evaluating agent-based systems and reasoning procedures in the context of automated driving. The aim here is to enable the model to independently generate further prompts and create action sequences. Through feedback loops and reflection, the model can correct itself, identify alternatives and monitor and improve its performance in new scenarios.

In the course of the project, the innovations described above will be integrated to validate the objectives in at least one vehicle demonstrator and evaluated on large data sets from the industrial partners. ADRIVE-GPT contributes with several innovations to the further development of highly scalable mobility technologies and strengthens Germany's position in the field of automated driving and artificial intelligence.