A Riemannian Framework for Learning Reduced-order Lagrangian Dynamics
Authors: Katharina Friedl, Noémie Jaquier, Jens Lundell, Tamim Asfour, Danica Kragic
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our approach by learning the dynamics of three simulated high-dimensional rigid and deformable systems: a pendulum, a rope, and a thin cloth. Our results demonstrate that it efficiently learns reduced-order dynamics leading to accurate long-term predictions of high-dimensional systems. ... 5 EXPERIMENTS ... Fig. 2-left shows the acceleration prediction errors on testing data for selected architectures trained on acceleration data. ... Table 1: Comparison of mean and standard deviation of prediction errors over 10 test trajectories. |
| Researcher Affiliation | Academia | 1 Division of Robotics, Perception, and Learning KTH Royal Institute of Technology ... 2 Institute for Anthropomatics and Robotics Karlsruhe Institute of Technology (KIT) |
| Pseudocode | No | The paper describes mathematical frameworks and network architectures with equations and flowcharts (e.g., Figure 1), but it does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor structured steps formatted like code. |
| Open Source Code | Yes | A video and source code are available at https://sites.google.com/view/reduced-lagrangians. |
| Open Datasets | No | We validate our approach by learning the dynamics of three simulated high-dimensional rigid and deformable systems: a pendulum, a rope, and a thin cloth. ... For this experiment, we use a 2-Do F pendulum implemented in MUJOCO (Todorov et al., 2012). ... The rope is implemented in MUJOCO via an elastic cable. ... The deformable thin cloth is modelled in MUJOCO as a flexible composite object. The paper describes simulated datasets generated using MUJOCO, but does not provide access information or explicitly state that these specific generated datasets are publicly available. |
| Dataset Splits | Yes | Each training dataset Dtrain consist of randomly-sampled data from 30 trajectories... The training datasets contain |Dtrain| = 8000 samples... Table 2 reports average reconstruction and prediction errors of RO-LNNs with latent dimensions d = {4, 6, 10, 14} on 10 testing trajectories. ... For training and testing, we generate datasets Dcloth = {{qn,k, qn,k, τn,k}K k=1}N n=1 consisting of N = 20 trajectories during training, and N = 10 trajectories during testing. |
| Hardware Specification | No | Averaged over 10 runs on the same local CPU, we achieve evaluation times of 113.59 s for the FOM, and 1.57 s for the ROM. The paper mentions experiments run on a 'local CPU' but does not specify the model or detailed specifications of the CPU. |
| Software Dependencies | No | We use a 2-Do F pendulum implemented in MUJOCO (Todorov et al., 2012). ... We use the Riemannian Adam (Becigneul & Ganea, 2019) implemented in Geoopt (Kochurov et al., 2020) to optimize the geometric LNN parameters. ... we implemented L-Op Inf in Python and solved the optimization problem leading to the reduced equations of motion with CVXPY. The paper mentions several software components (MUJOCO, Geoopt, Riemannian Adam, Python, CVXPY) but does not provide specific version numbers for any of them. |
| Experiment Setup | Yes | The models are trained by minimizing the LNN loss (6) for 3000 epochs with architecture-specific learning rates. ... The training datasets contain |Dtrain| = 8000 samples and are trained for 2000 epochs. ... For the first model, we use a learning rate of 5 10 2 for the AE parameters Ξ, 1 10 5 for the LNN parameters θ, and a regularization γ = 1 10 6. ... Both models are trained until convergence, i.e., for 4000 and 3000 epochs, respectively. |