Item active pick-up method through mechanical arm based on deep and reinforced learning

Abstract

The invention provides an item active pick-up method through a mechanical arm based on deep and reinforced learning, and belongs to the field of artificial intelligence and robotics applications. Themethod comprises the following steps of firstly establishing a simulation environment containing the machine arm and an item pick-up scene; establishing a deep Q learning network NQ based on multipleparallel U-Nets; carrying out multiple robot grabbing action policy tests in the simulation environment to perform training on the NQ, thus obtaining a trained deep learning network; establishing an item pick-up system in actual pick-up use, utilizing the trained deep learning network, inputting deep color images, and determining to utilize an action policy of actively changing a scene or pick upan item through the mechanical arm according to a measurement defined on a confidence image. According to the item active pick-up method disclosed by the invention, by actively changing an item pick-up environment through the mechanical arm and adapting to different pick-up conditions, a high success rate of pick-up can be achieved.

Description

technical field [0001] The invention belongs to the technical field of robot application, and in particular relates to a method for actively picking up items by a mechanical arm based on deep reinforcement learning. Background technique [0002] Robust and efficient item picking is one of the main research contents of robotics. With the rapid development of e-commerce, it is widely used in warehouse management, unmanned stores and industrial production lines. Most of the current item picking methods are based on passive methods, using cameras to capture static images of the current item stacking scene for item segmentation and classification and pose estimation. However, in practical applications, the item picking scene is complex, and it is difficult to perform accurate and efficient item segmentation. In classification and pose estimation, there are often phenomena that are not conducive to picking, such as mutual occlusion of items, and the approach of item poses to limit...

AI Technical Summary

Problems solved by technology

[0003] Researchers from MIT and Princeton published a paper entitled Robotic Pick-and-Place of Novel Objects in Clutter with Multi-Affordance Grasping and Cross-Domain Image Matching (through multi-confidence map capture and Cross-domain image matching enables robots to pick and place unfamiliar objects in stacking scenes). This technology is trained through deep learning and proposes the SuctionAffordances network (suction position confidence network), which inputs the color depth image of the picked scene and outputs pixel-level Suction Affordances Map (suction location confidence map), avoids complex item segmentation and identification, and directly obtains alternative pickup locations. However, in complex scenarios, the network is prone to output wrong values, resulting in pickup failure. Existing methods cannot effectively solve this problem. the problem

Images