The Technical University of Munich (TUM) carried out a research project funded by VDI/VDE Innovation + Technik GmbH under the grant number 16ES1029. The central scientific question addressed by the team was how to distribute computationally intensive artificial‑intelligence (AI) algorithms across the limited hardware resources of an autonomous vehicle in a way that satisfies multiple optimisation goals. These goals include minimising power consumption while guaranteeing that system performance does not fall below a specified threshold. Because driving situations change dynamically, the allocation of AI tasks must be re‑scheduled at run‑time, activating, deactivating or reallocating algorithms according to their criticality. The problem is formally modelled as a scheduling optimisation problem, a well‑known class in theoretical computer science.

To solve this, TUM first defined a meta‑model that captures both the characteristics of the embedded hardware (e.g., processing capacity, memory limits, communication bandwidth) and the requirements of the AI algorithms (e.g., execution period, deadline, safety criticality). From this meta‑model the team derived scheduling constraints that can be encoded in the language of solver technologies. The model‑based approach allowed the automatic generation of constraints that reflect the hardware capabilities and the AI workload. The derived constraints were then fed into optimisation solvers to compute schedules that meet the defined objectives.

The project was organised into a series of work packages. The initial packages focused on requirements analysis, architecture specification and the development of a first concept. Subsequent packages built a test environment and defined test scenarios that validate the environment’s behaviour. Dummy modules were integrated into the testbed, and the optimal resource‑scheduler concepts were implemented and tested against the defined requirements. A software framework was then planned and integrated into the target system, bringing the AI algorithms and the scheduler together. The final work packages involved presenting the scheduler’s runtime behaviour under different driving scenarios, performing comprehensive system‑level testing, debugging, and evaluating the overall integration. A demonstration system was used to showcase the scheduler’s performance and to document the demonstration process.

Performance evaluation results are presented in a series of figures. In a scenario with eight electronic control units (ECUs) and eight applications, each application consisting of two threads, the solution time for generating schedules grows exponentially as the number of threads per application increases. For example, when the number of applications is increased to 70 and distributed over eight ECUs, each ECU must schedule at least eight applications or 16 threads, leading to a dramatic increase in solution time. In a more complex zonal architecture with 15 ECUs and 60 safety‑critical applications, the scheduler was able to compute correct schedules, but the generation time again exhibited exponential growth due to the increased number of threads and the use of odd‑valued periods, which complicate hyper‑period calculations. Multi‑objective optimisation was also performed, taking into account end‑to‑end latency, response time, loss‑of‑response (LOR), resource utilisation and maximum memory usage, demonstrating the scheduler’s ability to balance competing objectives.

The project’s outcomes were made available to the wider community through an open‑source release of the software framework on GitHub (https://github.com/hadiaskaripoor/ee‑designer). This decision supports reproducibility, encourages industrial and academic collaboration, and provides a foundation that can be adapted to future projects. The collaboration involved TUM as the primary research partner, with VDI/VDE Innovation + Technik GmbH acting as the project sponsor and funder. The project spanned multiple phases from requirement analysis to final system integration, culminating in a demonstrator that showcased the scheduler’s capabilities in realistic autonomous driving scenarios.