Robots Can Learn to Grasp Objects Better with Deep Learning

Here are how robots can learn to grasp objects with the help of deep learning

Most adult individuals are born with the ability to pick up and handle things in their surroundings in ways that assist their utilization. When picking up a culinary tool, for example, they generally grasp it from the side that will not be placed into the cooking pot or pan.

Robots, on the other hand, must be taught how to pick up and handle items while doing various activities. This can be a difficult task since the robot may come across stuff it has never seen before.

The Autonomous Intelligent Systems (AIS) research group at the University of Bonn recently created a novel learning pipeline to increase a robotic arm's capacity to manage things in ways that better support their practical application. Their technique is done with the help of deep learning. This is described in a study published on the pre-print service arXiv, which might aid in the creation of robotic helpers capable of doing manual chores more successfully.

"An object is grasped functionally if it can be used, for example: an index finger on the trigger of a drill," one of the study's researchers, Dmytro Pavlichenko, told Tech Xplore. "Such a specific grasp may not always be reachable, necessitating manipulation." We discuss dexterous pre-grasp manipulation using an anthropomorphic hand in this study."

Pavlichenko and co-author Sven Behnke's current article builds on the AIS group's prior research efforts, namely a paper presented at the 2019 IEEE-RAS International Conference on Humanoid Robots in Toronto. The researchers devised a complicated technique for dual-arm robotic re-grasping of objects as part of their previous work, which depended on numerous intricate hand-designed components.

"The motivation for our new paper was to replace such a complex pipeline with a neural network," noted Pavlichenko. "This reduces complexity and removes hardcoded manipulation strategies, increasing the flexibility of the approach."

The researchers' new streamlined pre-grasp manipulation strategy is based on deep reinforcement learning, a high-performing and well-known technique for training AI systems. Using this method, the scientists taught a model to move things dexterously before gripping them, guaranteeing that the robot eventually holds them in effective ways, just as required.

"Our model learns by utilising a multi-component dense reward function, which incentivizes bringing an object closer to the given target functional grasp through finger-object interaction," Pavlichenko explained. "Learning can be done quickly when combined with a GPU-based simulation Isaac Gym."

So far, the researchers have tested their technique in a simulation environment called Isaac Gym and discovered that it yielded very promising results. In preliminary experiments, their methodology enabled virtual robots to learn how to move uniquely shaped items in their hands, finally determining the optimum technique to manage them without the need for human demonstrations.

Notably, the learning technique described by Pavlichenko and his Behnke may be readily extended to a wide range of robotic arms and hands, while also allowing for the manipulation of a wide range of items with varying forms. It might thus be deployed and evaluated on numerous physical robots in the future.