Zero-Shot Transfer of Neural ODEs

Authors: Tyler Ingebrand, Adam Thorpe, Ufuk Topcu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate state-of-the-art system modeling accuracy for two Mu Jo Co robot environments and show that the learned models can be used for more efficient MPC control of a quadrotor. (...) We demonstrate the effectiveness of our approach for predicting and controlling dynamical systems through several numerical experiments.
Researcher Affiliation Academia Tyler Ingebrand, Adam J. Thorpe, Ufuk Topcu University of Texas at Austin Austin, TX 78712
Pseudocode Yes Algorithm 1 Training Function Encoders with Neural ODE Basis Functions (...) Algorithm 2 The Residuals Method
Open Source Code Yes The source code is available at https://github.com/tyler-ingebrand/Neural ODEFunction Encoder. (...) We provide code as a zip file in the initial submission. A link to a github repository will be provided in the final version.
Open Datasets Yes We evaluate the performance of our proposed approach on the Half-Cheetah and Ant environments [28] (...) We use a simulated quadrotor system using Py Bullet [31]
Dataset Splits Yes Evaluations are done on a holdout set collected through the same means. (...) Evaluation is over 5 seeds, shaded regions show the first and third quartiles around the median. (...) Shaded region is 1st and 3rd quartiles over 200 trajectories (left) and over 5 trajectories (middle, right).
Hardware Specification Yes All experiments use an Intel 9th Generation i9 CPU and a Nvidia 2060 GPU with 6GB of memory.
Software Dependencies No The paper mentions software components like 'ADAM optimizer' and 'RK4 integrator' but does not specify version numbers for these or other key software libraries and dependencies.
Experiment Setup Yes We use an ADAM optimizer with a learning rate of 1e 3, and gradient clipping with a max norm of 1. NODE baselines uses 4 hidden layers of size 512, while FE + NODE baselines uses 4 hidden layers of size 51 for each basis function. All baselines train on 50 functions per gradient update via gradient accumulation. States are normalized to have 0 mean and unit variance.