On Foundation Models for Dynamical Systems from Purely Synthetic Data

TMLR Paper5767 Authors

Abstract:

Foundation models have demonstrated remarkable generalization, data efficiency, and robustness properties across various domains.
The success of these models is enabled by large-scale pretraining on Internet-scale datasets. These are available in fields like natural language processing and computer vision, but do not exist for dynamical systems. In this paper, we explore whether these properties can be achieved in the control domain when pretraining is performed entirely on synthetic data. We address this challenge by pretraining a transformer-based foundation model exclusively on synthetic data and propose to sample dynamics functions from a reproducing kernel Hilbert space. Our model, pretrained on purely synthetic data, generalizes to prediction tasks across different dynamical systems, which we validate in simulation and hardware experiments, including cart-pole and Furuta pendulum setups. Additionally, our model can be fine-tuned effectively to new systems increasing its performance even further. Our results demonstrate that even when pretrained solely on appropriately designed synthetic data, it is feasible to obtain foundation models for dynamical systems that outperform specialist models in terms of generalization, data efficiency, and robustness.

Submission Length: Regular submission (no more than 12 pages of main content)

Previous TMLR Submission Url: https://openreview.net/forum?id=gWsLNba94Q

Changes Since Last Submission: Template font has been corrected (unmodified).

Assigned Action Editor: ~Gilles_Louppe1

Submission Number: 5767