Graph Neural Networks for Robotic Grasping

Key information:

Abstract

Robotic systems provide the potential to support humans for physically strenuous tasks and in scenarios which are unsafe, such as search-and-rescue. To operate in such environments a robot needs to control how it physically interacts with its surroundings; the force it exerts on an object depends on the object’s properties such as weight and brittleness. Robots are equipped with sensors to recognise objects and infer physical properties. However, sensors or the object itself may be damaged, and robots cannot rely on haptics in the same way as humans.

The aim of this project is to allow robots to recognise objects and navigate environments using Graph Neural Networks (GNNs) and Few-shot learning. Learning from fewer examples is one of the big challenges currently facing machine learning. Suitable datasets for the task are unavailable for training, and GNNs are highly suited to Few-shot learning due to inferring missing labels from a graph which encodes the similarity between datapoints.

Many well-known machine learning frameworks can be thought of as special cases of GNNs. Reducing the computational cost of GNNs for real-world graphs is key for desirable embedded applications such as robotics. The ubiquity of graph-structured data means that progress in graph learning can facilitate breakthroughs across disciplines.