Advances in flexible manipulation through the application of AI-based techniques

The manufacturing industry was pioneer to incorporate robots to automate costly processes, with the aim of increasing production and reducing costs. Today, robots have a significant presence in almost every industry, generally carrying out simple and repetitive tasks. We are in the transition between Industry 4.0 and 5.0, where in addition to productivity, flexibility is also sought to adjust processes to specific customer needs.

Pick and place are two basic operations in practically any robotic application. Conventional robotic pick and place solutions currently in use in industry are characterised by their efficiency in performing simple, repetitive tasks. However, these are very rigid systems, work in completely controlled environments, and are very costly to reprogramme for other tasks. 

There are currently tasks in various industrial environments (e.g. order generation in a logistics environment) which require flexible handling of objects and which have not yet been automated due to their nature. The main bottlenecks which hinder their automation are the variety of objects to be handled, the lack of robot dexterity and the uncertainty introduced by uncontrolled dynamic environments.

In contrast, advanced robotics seeks system flexibility as well as productivity. This type of robotics is applied to problems that require advanced decision-making or unstructured environments, using collaborative robots. Artificial intelligence (AI) is playing an increasingly important role in robotics, as it provides robots with the necessary intelligence to solve complex tasks. In addition, AI allows complex behaviours to be learned using real-world experience, thus considerably reducing the cost of programming.

In view of the limitations of current robotic object manipulation systems, the main goal of this work is to increase the flexibility of manipulation systems using AI-based algorithms, providing robotic systems with the necessary capabilities to adjust to dynamic environments without the need for reprogramming. We focus on overcoming the limits in three specific lines of research related to pick and place:

• The first line of research focuses on studying the feasibility of learning a positioning policy for the base of a mobile manipulator, in order to enable the arm to pick up an object. Specifically, we propose to learn such behaviour using the experience acquired through interaction with the environment and thus avoiding programming it.
• The second line of research is focused on developing a grasping point detection system for arbitrary objects. The idea is that such a module will be able to estimate grasping points on a wide variety of objects in bin-picking scenes, thus avoiding the need to configure the system for each reference. Specifically, we propose to directly perform learning on n-dimensional point clouds labelled by an expert, and show that a higher accuracy is obtained than a 2D image-based reference system. 
• The last line of research is based on the development of a dynamic algorithm for the generation of object packing mosaics. Precisely, the proposed algorithm is able to compute mosaics with arbitrary objects in an online way. The validation of the system carried out on a real robotic prototype demonstrates the flexibility of the packaging system to generate mosaics with a wide variety of objects.