Geometric Contact Flows: Contactomorphisms for Dynamics and Control

Authors: Andrea Testa, Søren Hauberg, Tamim Asfour, Leonel Rozo

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on learning dynamics for physical systems and for controlling robots on interaction tasks demonstrate the effectiveness of our approach. ... 6. Experiments
Researcher Affiliation Collaboration 1Bosch Center for Artificial Intelligence, Renningen, Germany 2Institute for Anthropomatics and Robotics, Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany 3Section of Cognitive Systems, Technical University of Denmark (DTU), Lyngby, Denmark.
Pseudocode Yes Algorithm 1 Training a Contactomorphism φr(T) ... Algorithm 2 Training the Ensemble {φrn(T)}n [1,N]
Open Source Code Yes A full video of the robot experiments, including adaptability to unseen obstacles, and the repository implementing the GCF framework are available at https://sites.google. com/view/geometric-contact-flows.
Open Datasets Yes We consider a 60-dimensional dataset (Otness et al., 2021) describing the dynamics of a 2D square grid of nodes connected by springs. ... handwriting dynamics reconstruction using two datasets (Lemme et al., 2015; Fabi et al., 2022)
Dataset Splits No The training dataset comprises 20 spring-mesh systems, each characterized by a distinct set of initial conditions. While training data is mentioned, explicit splits for training/validation/testing with percentages or sample counts are not provided for any dataset.
Hardware Specification Yes The framework runs on a machine equipped with 13th Gen Intel Core i7-13850HX CPUs.
Software Dependencies No The paper mentions 'ROS2 acting as the middleware' but does not specify a version number. No other specific software components with version numbers are provided.
Experiment Setup Yes The weights employed in the loss function (12) are wx = 1 and wz = 0.01. Training is conducted for 5000 epochs, taking approximately 4 hours on average. ... The initial learning rate is 1 10 3 and is reduced by a factor of 0.9 on plateaus observed for 200 epochs. The loss is clipped at 1 103, and the gradient is clipped at 0.1. Optimization is performed using the Adam optimizer with default hyperparameters, and L2 regularization is applied to the weights with a coefficient of 1 10 10.