Overview
Learning control policies directly on physical systems—robots, vehicles, or industrial machines—is difficult because data collection is expensive, noisy, and safety-critical. Classical control theory provides strong guarantees but often requires accurate models; reinforcement learning can learn from data but typically needs large amounts of interaction and may be unsafe during exploration.
Led by Sebastian Trimpe at the Max Planck Institute for Intelligent Systems, the project “MachineData: Machine Data Sets and Their Use for Learning Control” (2018–2024) investigated how to design, collect, and exploit high-quality machine datasets for learning-based control.
Key results
- Developed methods for safe exploration in learning-based control, ensuring that learning algorithms do not violate safety constraints while collecting data on real hardware.
- Proposed GoSafeOpt, a scalable safe-exploration algorithm that can discover globally optimal policies for high-dimensional dynamical systems while providing safety and optimality guarantees.
- Created high-fidelity datasets for benchmark platforms such as the Mini Wheelbot and Crazyflie drones, including synchronized sensor readings, state estimates, motion-capture ground truth, and video logs across multiple controllers and environments.
- Advanced model-based reinforcement learning and Gaussian-process-based control, combining probabilistic models with optimal control to learn reliable policies from limited data.
- Developed event-triggered control and joint optimization approaches that learn when and how to update control actions from data, reducing communication and computation costs.
- Produced open datasets, software tools, and benchmark protocols to support reproducible research in learning control, alongside peer-reviewed publications on safe learning, model-based RL, and data-driven control.
Why it matters
The project addresses a central bottleneck in learning control: the need for large, diverse, and safe datasets from real machines. By providing methods for safe exploration and high-quality datasets, it enables more reliable, data-efficient learning of controllers for robots, drones, and other cyber-physical systems. The results can support applications in autonomous systems, industrial automation, and robotics, where safety and efficiency are critical.
This project produced the following peer-reviewed publications:
- Gräfe, A., Scherer, C., Hönig, W., & Trimpe, S. (2026). How to Model Your Crazyflie Brushless. arXiv:2603.05944 [cs.RO].
- Sukhija, B., Turchetta, M., Lindner, D., Krause, A., Trimpe, S., & Baumann, D. (2022). GoSafeOpt: Scalable Safe Exploration for Global Optimization of Dynamical Systems. arXiv:2201.09562 [cs.LG].
- Funk, N., Baumann, D., Berenz, V., & Trimpe, S. (2021). Learning Event-Triggered Control from Data through Joint Optimization. IFAC Journal of Systems and Control, 16, 100144. doi:10.1016/j.ifacsc.2021.100144
- Rohr, A., Neumann-Brosig, M., & Trimpe, S. (2021). Probabilistic Robust Linear Quadratic Regulators with Gaussian Processes. In Proceedings of the 3rd Conference on Learning for Dynamics and Control (L4DC 2021). Proceedings of Machine Learning Research, 144, 324–335.
- Hose, H., Brunzema, P., Subhasish, D., & Trimpe, S. (2025). The Mini Wheelbot Dataset: High-Fidelity Data for Robot Learning. arXiv:2601.11394 [cs.RO].
