01.06.2026
Manuel recently attended the IEEE ICRA 2026 conference in Vienna, Austria, where he presented our latest research findings to an international audience of researchers and experts.
At the conference, Manuel presented his paper “Lagrangian Neural Network-Based Control: Improving Robotic Trajectory Tracking via Linearized Feedback”, which was co-authored with Alexander Pawluchin, Jan-Hendrik Ewering, Thomas Seel, and Ivo Boblan. The paper was published in IEEE Robotics and Automation Letters (RA-L) and presented at ICRA as a poster, providing an opportunity to discuss the research findings with conference attendees.
The paper introduces LNN-CTC, a control framework that embeds a Lagrangian neural network (LNN) inside a computed torque controller. The LNN learns the robot's inverse dynamics directly from data while respecting Lagrangian mechanics, so no prior model of the robot is needed. Because the learned model sits inside the feedback-linearization loop instead of acting as a feedforward correction, it shapes the closed-loop error dynamics. On a robotic arm with four degrees of freedom, LNN-CTC trained on only 100 seconds of data reduced the tracking error by about a third compared with model-based computed torque control. It also outperformed PINN- and DNN-based computed torque control, an LNN feedforward controller, and a PID controller.
More background on the work is in the research project Lagrangian Neural Networks for Computed Torque Control.
Beyond the poster session, the conference provided an excellent opportunity for interesting discussions and conversations with researchers from various disciplines.
