Learning-based approaches for robotic grasping using visual sensors typically require collecting a large size dataset, either manually labeled or by many trial and errors of a robotic manipulator in the real or simulated world. We propose a simpler learning-from-demonstration approach that is able to detect the object to grasp from merely a single demonstration using a convolutional neural network we call GraspNet. In order to increase robustness and decrease the training time even further, we leverage data from previous demonstrations to quickly fine-tune a GrapNet for each new demonstration. We present some preliminary results on a grasping experiment with the Franka Panda cobot for which we can train a GraspNet with only hundreds of train iterations.
The Indian Energy grid is growing rapidly and there is a large simulation to improve not only the grid reliability, but also provide power for all by 2027. To this aim the Government of India has launched the Restructured Accelerated Power Development Program (RAPDRP). In India, residential loads contribute to about 27% of the total power consumption, and hence understanding the residential power consumption is a key stepping stone to improving the performance of the grid and optimizing energy usage via dynamic load management like demand response programs. While there is a lot of prior-art related to short and long term forecasting residential loads, similar models are still very limited in the Indian context. The Indian power grid is a complex eco-system, where demand and consumption patterns are influenced by weather, socio-economic status, geographic location, political events to name a few. Hence, it is difficult to build a model for energy forecasting at residential level. This paper is an attempt to build a residential forecasting model depending on the socio-economic factor and geographical location. To this aim, in this paper we first discuss the influence of socio-economic status on the energy consumption in urban and rural area’s. The socio-economic status includes number of earning members of the family, age, gender and profession etc. All these factors influence the number of devices within a home, which in turn impacts the overall energy consumption. Based on some of these observations, recommendation on modeling approaches are made to enable future data driven research into the energy consumption at residential level in India. This can further be used to enable smart grid technologies, especially from a smart demand, response perspective.
In this paper we discuss some issues in the experimental evaluation of intelligent autonomous systems, focusing on systems, like autonomous robots, operating in physical environments. We argue that one of the weaknesses of current experimental practices is the low degree of generalization of experimental results, meaning that knowing the performance a robot system obtains in a test setting does not provide much information about the performance the same system could achieve in other settings. We claim that one of the main obstacles to achieve generalization of experimental results in autonomous robotics is the low degree of representativeness of the selected experimental settings. We survey and discuss the degree of representativeness of experimental settings used in a significant sample of current research and we propose some strategies to overcome the emerging limitations.
Detection and tracking of moving objects with camera systems mounted on a mobile robot presents a formidable problem since the ego-motion of the robot and the moving objects jointly form a challengingly discernible motion in the image. In this paper, we are concerned with multiple-camera systems, namely the Ladybug\(^{\textregistered }2\) camera, whose perspective images were used to detect motion and subsequently perform the tracking of multiple objects on the sphere. This enabled us to account for the continuity of the scene which is achieved by the sensor in an image stitching process on the sphere. The objects are tracked on the sphere with a Bayesian filter based on the von Mises–Fisher distribution and the data association is achieved by the global nearest neighbor method, for which the distance matrix is constructed by deriving the Rényi \(\alpha \)-divergence for the von Mises–Fisher distribution. The prospects of the method are tested on a synthetic and real-world data experiments.
. Indoor localization and navigation systems for individuals with visual impairments (VI) typically rely upon extensive augmentation of the physical space or expensive sensors. Thus, few systems have been adopted. This work conducts a feasibility study of whether it is possible to localize and guide people with VI using inexpensive sensors, such as compasses and pedometers, which are available in portable devices like smart phones. The proposed approach takes advantage of interaction between the system and the human user, who confirms the presence of landmarks. Experiments are employed to study what kind of directions are successful in assisting human users to reach their destination. These experiments show that Bayesian localization tools provide sufficient accuracy, while achieving real-time operation, despite the minimalistic, noisy nature of sensors and the limited computational resources available on smart phones.
This work demonstrates how an autonomous robotic platform can use intrinsically noisy, coarse-scale visual methods lacking range information to produce good estimates of the location of objects, by using a map-space representation for weighting together multiple observations from different vantage points. As the robot moves through the environment it acquires visual images which are processed by means of a fast but noisy visual detection algorithm that gives bearing only information. The results from the detection are projected from image space into map space, where data from multiple viewpoints can intrinsically combine to yield an increasingly accurate picture of the location of objects. This method has been implemented and shown to work for object localization on a real robot. It has also been tested extensively in simulation, with systematically varied false positive and false negative detection rates. The results demonstrate that this is a viable method for object localization, even under a wide range of sensor uncertainties.
The vision of humans and robots working together as peers to accomplish complex tasks has motivated many recent research endeavors with a variety of applications ranging from lunar construction to soccer. However, much of this research is still at an early stage, and many challenges still remain in realizing this vision. A key requirement for enabling robustness and efficiency in human-robot teams is the ability to dynamically adjust the level of autonomy to optimize the use of resources and capabilities as conditions evolve. While sliding autonomy is well defined and understood in applications where a single human is working with a single robot, it is largely unexplored when applied to teams of humans working with multiple robots. This paper highlights the challenges of enabling sliding autonomy in peer-to-peer human-robot teams and extends the current literature to identify and extend six key capabilities that are essential for overcoming these challenges. These capabilities are requesting help, maintaining coordination, establishing situational awareness, enabling interactions at different levels of granularity, prioritizing team members, and learning from interactions. We demonstrate the importance of several of these characteristics with results from a peer-to-peer human-robot team engaged in a treasure hunt task.
. This paper proposes a novel way of characterizing the local geometry of 3D points, using persistent feature histograms. The relationships between the neighbors of a point are analyzed and the resulted values are stored in a 16-bin histogram. The histograms are pose and point cloud density invariant and cope well with noisy datasets. We show that geometric primitives have unique signatures in this feature space, preserved even in the presence of additive noise. To extract a compact subset of points which characterizes a point cloud dataset, we perform an in-depth analysis of all point feature histograms using different distance metrics. Preliminary results show that point clouds can be roughly segmented based on the uniqueness of geometric primitives feature histograms. We validate our approach on datasets acquired from laser sensors in indoor (kitchen) environments.
Focusing the development of non-industrial robotics in the last decade the growing impact of service and entertainment robots for daily life has emerged from pure science fiction to a serious scientific subject. But still many questions in how to solve everyday tasks like laying the table or even "simpler" detecting objects in unstructured areas with varying lighting conditions are unsolved. Hence the strong need to evaluate and exchange different approaches and abilities of multiple robotic demonstrators under real world conditions is also a crucial aspect in the development of system architectures. In this paper an architecture will be described providing strong support for simple exchange and integration of new robot abilities.
We present an implementation of the SRG method, a decentralized cooperative exploration strategy. In this method, a roadmap of the explored area with the associate safe region is stored in a compact data structure, called Sensor-based Random Graph (SRG). No task decomposition and/or allocation is performed. The roadmap is incrementally built through a simple decentralized mechanism: each robot moves towards its local frontier to access areas that, on the basis of the available information, appear to be unexplored by the rest of the team. A detailed description of the software architecture used to implement the strategy is given. Preliminary experiments with a team of Khepera III robots are presented to show the performance of the proposed technique.
We present an autonomous multi-robot system that can collect objects from indoor environments and load them into a dishwasher rack. We discuss each component of the system in detail and highlight the perception, navigation, and manipulation algorithms employed. We present results from several public demonstrations, including one in which the system was run for several hours and interacted with several hundred people.
Distributed heterogeneous robotic systems are often organized in component-based software architectures. The strong added value of these systems comes from their potential ability to dynamically self-configure the interactions of their components, in order to adapt to new tasks and unforeseen situations. However, no satisfactory solutions exist to the problem of automatic self-configuration. We propose a self-configuration mechanism where a special component generates, establishes and monitors the system configurations. We illustrate our approach on a distributed robotic system, and show an experiment in which the configuration component dynamically changes the configuration in response to a component failure.
This paper describes an autonomous learning method used with real robots in order to acquire ball passing skills in the RoboCup standard platform league. These skills involve precisely moving and stopping a ball to a certain objective area and are essential to realizing sophisticated cooperative strategy. Moreover, we propose a hybrid method using "thinning-out- and "surrogate functions" in order to reduce actual trials regarded as unnecessary or unpromising. We verify the performance of our method using the minimization problems of several test functions, and then we address the learning problem of ball passing skills on real robots, which is also the first application of thinning-out on real environments.
The paper considers the problem of synthesising accurate dynamic models of a miniature rotorcraft based on minimal physical assumptions, and using the models to develop a controller. The approach is based on the idea of building models that predict accelerations, and is implemented using evolutionary programming in a particularly efficient co-evolutionary framework. Both the structure and the parameters of the nonlinear models are jointly identified from real data. The modelling method is demonstrated and validated on a miniature quadrotor rotorcraft, and a controller based on the model is then developed and tested.
In this paper, we propose a method to estimate a robot position and orientation in a room by using a map of color histograms. The map of color histograms is a distribution map of colors measured by a mobile robot in multiple directions at various points. The size of the map should be strongly compressed because the memory capacity of the robot is restricted. Thus, the histograms are converted by the discrete cosine transform, and their high-frequency components are cut. Then, residual low-frequency components are stored in the map of color histograms. Robot localization is performed by matching between the histogram of an obtained image and reconstructed histograms out of the map by means of histogram intersection. We also introduce a series of estimations with the aid of odometry data, which leads to accuracy improvement.
Analysis of continuous process of motor learning gives a lot of useful knowledge for the recovery of human motion activities, and functional adaptability in new environment. This paper proposes a valuation index for degree of proficiency, and shows results of motor skill analysis for several arm reaching tasks. The motor skills were evaluated by using the reproducibility of muscle activation patterns, which were represented by using the variance value of the Electromyographic (EMG) signal patterns, and the motion accuracy. We confirm that the reproducibility is high when the motion accuracy is high, and the various skill acquisition processes exist due to individual difference. We conclude that, the reproducibility is one of the important indices for evaluating the degree of proficiency.
This paper describes a competitive learning method for a robust EMG-to-motion classifier system, to adjust to the change in user's characteristics. This method is done under the assumptions that the input motions are continuous, and the teaching motions are ambiguous in nature, therefore, automatic addition, elimination and selection of learning data are possible. Applying our proposed method, we conducted experiments to discriminate eight forearm motions, with the results, a stable and highly effective discrimination rate was achieved and maintained even when changes occurred in user's characteristics.
One of the most graceful phenomena widely observed in nature is self-assembly; living systems spontaneously form their body structure through the developmental process. While this remarkable phenomenon still leaves much to be understood in biology, the concept of self-assembly becomes undeniably indispensable also in artificial systems as they increase in size and complexity. Based on this consideration, this paper discusses the realization of self-assembly with the use of a modular robot. The main contributions of this paper are twofold: the first concerns the exploitation of emergent phenomena stemming from the interplay between the control and mechanical systems; the second is related to the implementation of different adhesiveness among the modules. Here, form generation by self-assembly is considered as the result of time evolution toward the most dynamically stable state. Preliminary simulation results show that stable and spontaneous self-assembly is achieved irrespective of the initial positional relationship among the modules.
Natural language communication is very important in Human-Robot cooperative systems. As a part of natural language communication, learning to identify natural language references such as "small red cube" is very useful. This paper proposes a method of learning object identification in which the meanings of references such as above can be inferred. The proposed method is demonstrated with an experiment conducted with a PA-10 redundant manipulator and a set of objects. The robot manipulator is issued commands like "pick the big red cube" to identify and pick objects placed on a table. The robot learns to interpret the meaning of this type of natural commands by learning individual lexical symbols in the vocabulary and their corresponding object features.
