Reyes Rios-Cabrera

Discriminatively Trained Templates for 3D Object Detection: A Real Time Scalable Approach

In this paper we propose a new method for detecting multiple specific 3D objects in real time. We start from the template-based approach based on the LINE2D/LINEMOD representation introduced recently by Hinterstoisser et al., yet extend it in two ways. First, we propose to learn the templates in a discriminative fashion. We show that this can be done online during the collection of the example images, in just a few milliseconds, and has a big impact on the accuracy of the detector. Second, we propose a scheme based on cascades that speeds up detection. Since detection of an object is fast, new objects can be added with very low cost, making our approach scale well. In our experiments, we easily handle 10-30 3D objects at frame rates above 10fps using a single CPU core. We outperform the state-of-the-art both in terms of speed as well as in terms of accuracy, as validated on 3 different datasets. This holds both when using monocular color images (with LINE2D) and when using RGBD images (with LINEMOD). Moreover, we propose a challenging new dataset made of 12 objects, for future competing methods on monocular color images.

Machine vision approach for robotic assembly

Purpose – Outcome with a novel methodology for online recognition and classification of pieces in robotic assembly tasks and its application into an intelligent manufacturing cell.Design/methodology/approach – The performance of industrial robots working in unstructured environments can be improved using visual perception and learning techniques. The object recognition is accomplished using an artificial neural network (ANN) architecture which receives a descriptive vector called CFD&POSE as the input. Experimental results were done within a manufacturing cell and assembly parts.Findings – Find this vector represents an innovative methodology for classification and identification of pieces in robotic tasks, obtaining fast recognition and pose estimation information in real time. The vector compresses 3D object data from assembly parts and it is invariant to scale, rotation and orientation, and it also supports a wide range of illumination levels.Research limitations/implications – Provides vision guidance...

Acquisition of welding skills in industrial robots

Purpose – The purpose of this paper was to propose a method based on an Artificial Neural Network and a real-time vision algorithm, to learn welding skills in industrial robotics. Design/methodology/approach – By using an optic camera to measure the bead geometry (width and height), the authors propose a real-time computer vision algorithm to extract training patterns and to enable an industrial robot to acquire and learn autonomously the welding skill. To test the approach, an industrial KUKA robot and a welding gas metal arc welding machine were used in a manufacturing cell. Findings – Several data analyses are described, showing empirically that industrial robots can acquire the skill even if the specific welding parameters are unknown. Research limitations/implications – The approach considers only stringer beads. Weave bead and bead penetration are not considered. Practical implications – With the proposed approach, it is possible to learn specific welding parameters despite of the material, type of ...

On-line incremental learning for unknown conditions during assembly operations with industrial robots

The assembly operation using industrial robots can be accomplished successfully in well-structured environments where the mating pair location is known in advance. However, in real-world scenarios there are uncertainties associated to sensing, control and modelling errors that make the assembly task very complex. In addition, there are also unmodeled uncertainties that have to be taken into account for an effective control algorithm to succeed. Among these uncertainties, it can be mentioned disturbances, backlash and aging of mechanisms. In this paper, a method to overcome the effect of those uncertainties based on the Fuzzy ARTMAP artificial neural network (ANN) to successfully accomplish the assembly task is proposed. Experimental work is reported using an industrial 6 DOF robot arm in conjunction with a vision system for part location and wrist force/torque sensing data for assembly. Force data is fed into an ANN evolving controller during a typical peg in hole (PIH) assembly operation. The controller uses an incremental learning mechanism that is solely guided by the sensed forces. In this article, two approaches are presented in order to compare the incremental learning capability of the manipulator. The first approach uses a primitive knowledge base (PKB) containing 16 primitive movements to learn online the first insertion. During assembly, the manipulator learns new patterns according to the learning criteria which turn the PKB into an enhanced knowledge base (EKB). During a second insertion the controller uses effectively the EKB and operation improves. The second approach employs minimum information (it contains only the assembly direction) and the process starts from scratch. After several operations, that knowledge base increases by including only the needed patterns to perform the insertion. Experimental results showed that the evolving controller is able to assemble the matting pairs enhancing its knowledge whenever it is needed depending on the part geometry and level of expertise. Our approach is demonstrated through several PIH operations with different tolerances and part geometry. As the robot’s expertise evolves, the PIH operation is carried out faster with shorter assembly trajectories.

On the design of a multimodal cognitive architecture for perceptual learning in industrial robots

Robots can be greatly benefited from the integration of artificial senses in order to adapt to changing worlds. To be effective in complex unstructured environments robots have to perceive the environment and adapt accordingly. In this paper, it is introduced a biology inspired multimodal architecture called M2ARTMAP which is based on the biological model of sensorial perception and has been designed to be a more versatile alternative to data fusion techniques and non-modular neural architectures. Besides the computational overload compared to FuzzyARTMAP, M2ARTMAP reaches similar performance. This paper reports the results found in simulated environments and also the observed results during assembly operations using an industrial robot provided with vision and force sensing capabilities.

Learning and fast object recognition in robot skill acquisition: a new method

Invariant object recognition aims at recognising an object independently of its position, scale and orientation. This is important in robot skill acquisition during grasping operations especially when working in unstructured environments. In this paper we present an approach to aid the learning of manipulative skills on-line. We introduce and approach based on an ANN for object learning and recognition using a descriptive vector built on recurrent patterns. Experimental learning results using a fast camera are presented. Some simple parts (i.e. circular, squared and radiused-square) were used for comparing different connectionist models (Backpropagation, Perceptron and FuzzyARTMAP) and to select the appropriate model. Later during experiments, complex figures were learned using the chosen FuzzyARTMAP algorithm showing a 93.8% overall efficiency and 100% recognition rate with not so complex parts. Recognition times were lower than 1 ms, which clearly indicates the suitability of the approach to be implemented in robotic real-world operations.

A learning approach for on line object recognition tasks

The performance of industrial robots working in unstructured environment can be improved using visual perception and learning techniques. In this work, a novel approach that uses 2D data and simple image processing techniques is introduced. A unique image vector descriptor (CFD&POSE) containing also depth information is computed and then input to a Fuzzy ART MAP architecture for learning and recognition purposes. This vector compresses 3D object data from assembly parts and is invariant to scale, rotation and orientation. The approach in combination with the fast learning capability of ART networks indicates the suitability for industrial robot applications as it is shown in experimental results.

A hybrid non-invasive method for internal/external quality assessment of potatoes

Consumers purchase fruits and vegetables based on its quality, which can be defined as a degree of excellence which is the result of a combination of characteristics, attributes and properties that have significance for market acceptability. In this paper, a novel hybrid active imaging methodology for potato quality inspection that uses an optical colour camera and an infrared thermal camera is presented. The methodology employs an artificial neural network (ANN) that uses quality data composed by two descriptors as input. The ANN works as a feature classifier so that its output is the potato quality grade. The input vector contains information related to external characteristics, such as shape, weight, length and width. Internal characteristics are also accounted for in the input vector in the form of excessive sugar content. The extra sugar content of the potato is an important problem for potato growers and potato chip manufacturers. Extra sugar content could result in diseases or wounds in the potato tuber. In general, potato tubers with low sugar content are considered as having a higher quality. The validation of the methodology was made through experimentation which consisted in fusing both, external and internal characteristics in the input vector to the ANN for an overall quality classification. Results using internal data as obtained from an infrared camera and fused with optical external parameters demonstrated the feasibility of the method since the prediction accuracy increased during potato grading.

Reconfigurable Distributed Controller for Welding and Assembly Robotic Systems: Issues and Experiments

Industrial production systems for smart factories or the so-called Industry 4.0 will demand high interoperability and connectivity between production modules, so that modules could be monitored in real-time. Production modules should make decisions on their own without human intervention; and they must be modular and adaptive to changing circumstances and customers’ requirements. The autonomous operation of production modules in smart factories imposes asynchronous delays due to several reasons, such as object recognition time, grasping time or welding delays that change due to a newly reoriented or positioned component. Consequently, production modules need to be speeded up to compensate for the delays in the previous production stages. In this paper, we present a novel Reconfigurable Distributed Controller (RDC) for Intelligent Robotic Welding and Assembly Systems that autonomously compensate the production delays. The proposed RDC compensates for three types of major production delays that affect the total production time. (I) The first delay can occur at individual level. In this case, the module can fully compensate, since no other modules are affected and the total production time for this product can be met. (II) The second type of delay occurs at inter-module level, where delays are so long that more than one production module will need to be reconfigured. (III) Finally, the third type of delay occurs in the worst-case scenario when the total production time cannot be met by modifying individual module’s production time. A total cell reconfiguration is needed, which implies to speed up the next production cycle to deliver the following product before its deadline. By doing so, the mean production time is maintained. In this paper, issues and experiments that show the feasibility of the RDC are presented. Results of using a distributed reconfigurable manufacturing cell composed of three industrial robots, conveyor belts, and a positioning table demonstrated the effectiveness of our approach to compensate the major delays in real working environments.

Proceso de Aprendizaje con Algoritmo Robusto para la Obtención del POSE de Objetos en Líneas de Ensamble con Robots en Tiempo Real (RT)

En el articulo, se presenta la metodologia y el algoritmo utilizado para obtener posicion y orientacion en tiempo real (POSE) de objetos que van a ser ensamblados en linea y en procesos de manufactura robotizados. La informacion del POSE y caracteristicas del objeto son integrados en un vector descriptivo que es utilizado por una red neuronal del tipo FuzzyARTMAP, para aprender del objeto y posteriormente reconocerlo y localizarlo. Se obtiene asi un proceso de aprendizaje de la locacion de piezas para su posterior reconocimiento y manipulacion por el robot manipulador. Se presenta la explicacion del algoritmo utilizado y su simulacion en MatLab 5.0, con casos concretos de formas basicas de piezas de ensamble. Se presentan tambien resultados experimentales dentro de una celda de manufactura con un robot de seis grados de libertad Kuka R15