The “Rob‑aKademI” project, funded by the German Federal Ministry of Education and Research under grant 01IS20009D, ran from 1 July 2020 to 30 June 2022 and was extended to 31 December 2022. Its goal was to create robust control algorithms and physics‑based simulation environments that enable industrial robots to be trained with artificial‑intelligence methods for highly flexible, variant‑rich assembly in “lot size 1”. The consortium comprised the Institute for Industrial Production and Factory Operation (IFF) at the University of Stuttgart, Fraunhofer Institute for Production Systems and Automation (IPA), and the companies TruPhysics GmbH, elprotek GmbH, Walter Meile GmbH, Dresen Elektronik, Kaepple Qualitätsleister, and micropsi industries. IFF led the project’s administrative and technical coordination (AP 1), the development of the Hybrid Machine Learning Framework (HMLF) (AP 3), the integration of modules into a training environment (AP 5), and the dissemination of results (AP 7). Fraunhofer IPA provided technical expertise and infrastructure support, while the industrial partners supplied real‑world use cases and validation scenarios.

The technical core of the project is the modular Hybrid Machine Learning Framework, which connects reinforcement‑learning (RL) algorithms to a simulation environment and the skill‑based robot‑control package Pitasc. HMLF was designed to be extensible, allowing new RL algorithms, simulation tools, and skill packages to be added with minimal effort. The framework incorporated four state‑of‑the‑art RL algorithms: Deep Deterministic Policy Gradient (DDPG), Soft Actor‑Critic (SAC), Twin‑Delayed DDPG (TD3), and Proximal Policy Optimization (PPO). These algorithms were trained on a Peg‑in‑Hole task performed by a UR5 industrial robot within a simulated environment. The task was split into a “Search” phase, where the robot locates the hole, and an “Insertion” phase, where the peg is inserted. Sensor data, such as the force applied to the table, were used as environmental parameters to provide realistic feedback. The modular pipeline enabled the separate training of the two phases and the seamless integration of the trained policies into the Pitasc skill set.

The project demonstrated that the HMLF pipeline could be deployed in a Docker container, ensuring reproducibility and ease of use for partner companies. The integration of HMLF, the simulation environment, Pitasc, and the RL algorithms was validated on the Peg‑in‑Hole use case, confirming that the framework meets the requirements for competency‑based training of industrial robots. The modular design allows the same pipeline to be applied to other assembly tasks, potentially reducing downtime and error rates in production lines. Although the report does not provide quantitative performance metrics, the successful training of the UR5 robot on the Peg‑in‑Hole task indicates that the framework achieves the intended learning objectives.

Beyond the technical achievements, the project emphasized collaboration and knowledge transfer. IFF managed intellectual‑property rights and data security, coordinated consortium meetings, and organized public dissemination through interviews, publications, and trade‑fair presentations. The extended six‑month period, funded by IFF, accommodated additional personnel to address unforeseen challenges and ensure the timely completion of the project’s milestones. Overall, Rob‑aKademI delivered a versatile, modular RL‑based training framework that bridges simulation and real‑world robot control, supporting the broader Industry 4.0 vision of flexible, digital‑twinned manufacturing systems.