Perceived Constraint Identification Using Physics Informed Deep Neural Networks

Autonomous manipulation for robots in unstructured environments is challenging due to the technical gap between perception and acting on the physical world. This challenge becomes even more difficult when an object's motion is constrained in space. This paper presents a method to rapidly estimate mechanical constraint models of systems in the environment using physics-informed deep neural networks (PINN). We develop a single network capable of estimating the motion of constrained mechanisms and present the methods for model training using synthetic data. By leveraging 6-dimensional constraints to selectively inform the loss function for different motions within the deep neural network, we achieve high accuracy in estimating the motions for linear and rotation constraints with sub-1 degree error, which is an order of magnitude improvement over recent studies. We experimentally evaluate our model on three real examples, validating the applicability of the approach.