05.10.2026
On 5 October 2026, Manuel presented a poster at Visions4AI at Leibniz Universität Hannover. Organized by the L3S AI research center, the event brings together science, business, politics, and society to discuss current developments in AI, with talks by international experts and poster and demonstrator presentations from researchers of Leibniz Universität Hannover.
Manuel's poster, "A Learning-Based Proprioceptive Pipeline: Sensorless Contact Detection and Classification for Human-Robot Collaboration", was co-authored with Jan Piosik, Moritz Schappler, and Thomas Seel. Safe human-robot collaboration requires robots to detect collisions rapidly and to distinguish them from intentional interactions such as hand-guiding. The presented pipeline does this using only proprioceptive signals, without external sensors, an analytically identified robot model, or labeled collision data.
In the first stage, a physics-structured Lagrangian neural network learns the robot's nominal energy from collision-free data, and an autoencoder (LNN-TAAE) flags deviations in the resulting energy residual as collisions. It detects collisions with a mean delay of 4.7 ms, the shortest of eight compared observers, at a mean F1-score of 92.86 % and 0.65 ms CPU inference time. In the second stage, a recurrent neural network classifies each detected contact on a planar parallel robot as intentional or unintentional, reaching 91.67 % online accuracy on 60 new contact events, ahead of a threshold-based baseline (85 %) and a Random Forest (80 %).
The poster session provided a great opportunity for lively discussions with researchers and visitors from a wide range of backgrounds.
