Context-Aware Deep Lagrangian Networks for Model Predictive Control

Lucas Schulze1, Jan Peters1, Oleg Arenz1
1Intelligent Autonomous Systems Lab, TU Darmstadt
Paper arXiv

Overview

The following highlights the main contributions of our work:

  • Combine Deep Lagrangian Networks (DeLaN) with a nominal model to predict only the unknown dynamics.
  • Extend DeLaN to a contextual setting with a history encoder, enabling a single network to adapt to different dynamics and perform online identification of physically plausible models.
  • Integrate the Context-Aware DeLaN into an MPC — CaDeLaC — built using software frameworks (e.g., PyTorch, Acados, Pinocchio, L4CasADi).
  • Demonstrate the performance of CaDeLaC in zero-shot experiments on a real 7-DOF Franka Emika Panda robot, using a model trained fully in simulation.

High-Speed Trajectories

CaDeLaC tracks high-speed trajectories near joint velocity limits with different payloads (1 kg, 2 kg, 3 kg), outperforming pure gravity compensation methods.

Pick-and-Place

Our method performs a pick-and-place task with 3 kg and 2 kg payloads, adapting online to changing dynamics in the environment.

BibTeX

If you find this work useful, please consider citing:

@misc{schulze2025contextawaredelan,
  title={Context-Aware Deep Lagrangian Networks for Model Predictive Control}, 
  author={Lucas Schulze and Jan Peters and Oleg Arenz},
  year={2025},
  eprint={2506.15249},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2506.15249} 
}

Acknowledgment

This project has been funded by the German Federal Ministry of Research, Technology and Space (BMBFTR) - Project number 01IS23057B.