- TL;DR: We propose a stochastic differentiable forward dynamics predictor that is able to sample multiple physically plausible trajectories under the same initial input state and show that it can be used to train model-free policies more efficiently.
- Abstract: Recently, neural-network based forward dynamics models have been proposed that attempt to learn the dynamics of physical systems in a deterministic way. While near-term motion can be predicted accurately, long-term predictions suffer from accumulating input and prediction errors which can lead to plausible but different trajectories that diverge from the ground truth. A system that predicts distributions of the future physical states for long time horizons based on its uncertainty is thus a promising solution. In this work, we introduce a novel robust Monte Carlo sampling based graph-convolutional dropout method that allows us to sample multiple plausible trajectories for an initial state given a neural-network based forward dynamics predictor. By introducing a new shape preservation loss and training our dynamics model recurrently, we stabilize long-term predictions. We show that our model’s long-term forward dynamics prediction errors on complicated physical interactions of rigid and deformable objects of various shapes are significantly lower than existing strong baselines. Lastly, we demonstrate how generating multiple trajectories with our Monte Carlo dropout method can be used to train model-free reinforcement learning agents faster and to better solutions on simple manipulation tasks.
- Keywords: physics prediction, forward dynamics, stochastic environments, dropout
- Original Pdf: pdf