Augmenting Differentiable Simulators with Neural Networks to Close the Sim2Real Gap
arXiv:2007.06045
Abstract
We present a differentiable simulation architecture for articulated rigid-body dynamics that enables the augmentation of analytical models with neural networks at any point of the computation. Through gradient-based optimization, identification of the simulation parameters and network weights is performed efficiently in preliminary experiments on a real-world dataset and in sim2sim transfer applications, while poor local optima are overcome through a random search approach.
References in corpus (10)
- Learning agile and dynamic motor skills for legged robots
- Interaction Networks for Learning about Objects, Relations and Physics
- Hamiltonian Neural Networks
- Learning to Simulate Complex Physics with Graph Networks
- Learning to Predict the Cosmological Structure Formation
- Deep Lagrangian Networks: Using Physics as Model Prior for Deep Learning
- Lagrangian Neural Networks
- Automatic Differentiation of Rigid Body Dynamics for Optimal Control and Estimation
- Interactive Differentiable Simulation
- Data-Augmented Contact Model for Rigid Body Simulation