Christian Dietz

Siemens Munich

Tuesday, December 09, 2025, 11:00 - 11:59

Building 102 - SR 01-012

Current model-based approaches for planning robotic assembly motions primarily rely on sampling-based methods or derivative-free reinforcement learning. The main difficulty for derivative-based methods is caused by the underlying contact dynamics, which inherently include several causes of nondifferentiability. The contact dynamics formulation can be split into two hierarchically structured subproblems, the collision detection and the contact resolution problem. We propose a dynamic formulation that smoothes both subproblems, motivated by the smoothing mechanism used in interior-point methods. Based on the smoothed dynamics, a multi-scenario-based optimal control formulation is utilized to compute robust assembly motions. In particular, a reference trajectory is determined, such that tracking it by a state feedback controller results in reliable execution of assembly motions on real robots. We carefully investigate the effect that dynamics smoothing and robust modelling has on the success rates in real-world experiments. Additionally, we show that if IPOPT is used to solve the considered optimal control formulation, using exact Hessians over commonly used approximations such as Gauss-Newton or L-BFGS results in improved convergence speed.