I’m Leo, a robot learning and AI engineer with an M.Sc. in Robotics, Cognition, and Intelligence from TUM. I’ve worked across the robotics and AI stack, with experience spanning language-guided dexterous manipulation for my master’s thesis with a robotics startup, neuromorphic computing at BMW Research, safe human-robot collaboration at TUM published in IEEE T-RO, and 3D perception/SLAM.
Today, I’m focused on foundation models for physical AI and exploring agentic AI through hackathons. Born in Norway and raised in Germany, I’m particularly excited to contribute to Europe’s deep-tech ecosystem and build ambitious robotics and AI systems here.
End-to-end visuomotor policy learning on a physical SO-101 manipulator. Built the full imitation-learning pipeline with LeRobot: teleoperation, dataset collection, ACT and SmolVLA training, and real-robot policy deployment.
Language-guided tool-motion generation for dexterous manipulation. Grounds natural-language instructions in parameterized SE(3) motion primitives, combining LLM-level reasoning with a learned low-level policy.
Neuro-inspired computing with Oscillatory Neural Networks in collaboration with IBM. Built Python/PyTorch simulators and collected tactile interaction data from 10 participants for a human-machine interaction demonstrator.
Runtime safety for physical human-robot interaction using reachability analysis and verification. Contributed to SaRA-Shield, providing formal safety guarantees for autonomous manipulation; published in IEEE Transactions on Robotics.
Browser-native interactive demos: a Unitree G1 locomotion simulator, 3D reconstruction with SLAM, and an RL-powered Pong agent.
Agentic AI product built at Hack-Nation’s Global AI Hackathon; invited to Venture Lab.