This is code for Adaptive Meta-Learning for Identification of Rover-Terrain Dynamics, in which we present a meta-learning based approach to adapt probabilistic predictions of rover dynamics and estimates of terrain parameters by augmenting a nominal model affine in parameters with a Bayesian regression algorithm (P-ALPaCA).
P-ALPaCA is an alternate formulation of ALPaCA, which is a framework for online learning that can be imbued with rich, informative priors offline to enable few-shot learning with Bayesian uncertainty estimates.
To use this codebase, first install the requirements by running the following line (ideally within a virtual environment)
pip install -r requirements.txt
Having done this, run the notebooks in the demos directory to train models, predict rover dynamics, estimate terrain parameters, and visualize the effect of orthogonality regularization.