Master Thesis Title: Physical Recursive Self Improvement (P-RSI)

What if a robot’s training pipeline could improve itself? Today, getting a reinforcement-learning policy from simulation onto a real robot takes long, manual tuning by a few experts. Recursive self-improvement (RSI) offers a potential solution by enabling systems to optimize not only their task policies, but also the underlying algorithms and environments that generate them. While RSI has achieved success in software domains through automated code generation and unit testing, extending these mechanisms to physical robotics introduces severe challenges, including partial observability, shifting dynamics, and the high cost of real-world failures. In this thesis, you explore how an AI system can take over parts of that loop and learn from conducting frequent sim-to-sim and periodic sim-to-real transfers. Instead of tuning the pipeline for a single task, the system turns lessons from sparse real-robot trials into general training rules. It then validates and refines these rules across diverse simulated scenarios, steadily improving its own training process with minimal physical risk.

Environment: The thesis is conducted at the RAI Institute, a top-notch robotics research institute with representation in Zurich. It would happen as part of a collaboration agreement between RAI and ETH Zurich. A mature robotics stack running on a multitude of robots and enabling fast testing cycles, strong engineering support, and a best-in-class robotics research laboratory with state-of-the-art computing infrastructures led by distinguished scientists in the field make this opportunity second to none. For examples of the platforms and deployments developed at the institute, please have a look at the videos on our channel
https://www.youtube.com/@rai-inst/videos