MI-HGNN: Morphology-Informed Heterogeneous Graph Neural Network for Legged Robot Contact Perception

TL;DR

Although the proposed Morphology-Informed Heterogeneous Graph Neural Network is applied to contact perception problems for legged robots in this work, it can be seamlessly applied to other types of multi-body dynamical systems and has the potential to improve other robot learning frameworks.

Abstract

We present a Morphology-Informed Heterogeneous Graph Neural Network (MI-HGNN) for learning-based contact perception. The architecture and connectivity of the MI-HGNN are constructed from the robot morphology, in which nodes and edges are robot joints and links, respectively. By incorporating the morphology-informed constraints into a neural network, we improve a learning-based approach using model-based knowledge. We apply the proposed MI-HGNN to two contact perception problems, and conduct extensive experiments using both real-world and simulated data collected using two quadruped robots. Our experiments demonstrate the superiority of our method in terms of effectiveness, generalization ability, model efficiency, and sample efficiency. Our MI-HGNN improved the performance of a state-of-the-art model that leverages robot morphological symmetry by 8.4% with only 0.21% of its parameters. Although MI-HGNN is applied to contact perception problems for legged robots in this work, it can be seamlessly applied to other types of multi-body dynamical systems and has the potential to improve other robot learning frameworks. Our code is made publicly available at https://github.com/lunarlab-gatech/Morphology-Informed-HGNN.

Figures (5)

• Figure 1: Visualization of our MI-HGNN for the Mini-Cheetah robot as an example. The structure and connectivity of our graph is constructed from the robot morphology. Local sensor measurements are embedded into the corresponding node to predict contact information at the foot node.
• Figure 2: Overview of the proposed MI-HGNN for legged robot contact perception problems. Our MI-HGNN is constructed from a robot kinematic structure where nodes are joints and edges are links. Proprioceptive sensor measurements acquired at each local frame are embedded into the corresponding node through a heterogeneous encoder, and fused via several layers of Message Passing. A foot decoder attached to the foot node exacts the contact information during inference.
• Figure 3: Contact detection results on the real-world Mini-Cheetah contact dataset lin_2021_CORL: (a) classification performance of four models on the unseen test set, trained with the entire training set. The mean and standard deviation across 8 runs are reported. (b) sample efficiency evaluation for all models.
• Figure 4: Examples of various terrains used for GRF data collection: top-left: flat terrain; bottom-left: 20$\degree$ slope; right: rough terrain. GRF for each leg in Z direction is visualized as the black arrow.
• Figure 5: Evaluation of all model types on a sub sequence of the "Unseen All" test sequence for the GRF estimation task, which includes an unseen friction coefficient ($\dot{x} = 0.5$), unseen speed ($v = 1.0$), and unseen terrain (rough).