encoder_dynamics: Anomaly detection for generic failure monitoring in robotic assembly, screwing and manipulation

Official code of the paper: Grambow, N., Fenner, L.-M., Kempkes, F., Hotz, P., Wan, D., Krüger, J., & Haninger, K. Anomaly detection for generic failure monitoring in robotic assembly, screwing and manipulation. IEEE Robotics and Automation Letters. https://doi.org/10.1109/LRA.2026.3664647 (preprint: arXiv:2509.26308).

The repository provides tools for time-series analysis of robot sensor data (e.g. force/torque signals) with autoencoder-based models. Models can be trained to reconstruct the sensor inputs, where the reconstruction loss is used to detect anomalies, or to predict future states for use in robot planning. Features include a multi-trainer for reconstruction and prediction, a JAX implementation with JIT compilation and GPU acceleration, and ROS/ROS 2 compatibility. Within CONVERGING, it is used as part of the PAM module for anomaly detection in robotic cabling.

Source code: https://gitlab.cc-asp.fraunhofer.de/ipk_aut/techmodules/encoder_dynamics/