Search

As humans, our grasping and manipulation skills are highly dependent on our ability to perceive tactile properties. Conversely, most of today's robotic operations still rely predominantly on visual feedback for identifying the objects that need to be grasped and handled. In this work, we study the problem of recognizing everyday objects based solely on their tactile attributes. This has a significant practical value, as it could allow object identification even when visual sensing is impossible, or assist vision in difficult contexts. Our method consists of acquiring multi-modal tactile sensing data during a quick grasp-centric exploration phase, with minimal operational cost. Our algorithm was able to recognize objects from a considerably-large set of 50 general purpose items with an accuracy of 98.1%. Moreover, we show that it is possible to reliably identify a large proportion of these objects by only analyzing the deformation pattern that they undergo during compression. Further, we study our method's ability to learn relevant tactile properties to classify new objects. We also share our tactile sensing database that contains various sensor data acquired from more than 1600 experiments, which was used for this work. Finally, we discuss the relative performance and role of each tactile modality for differentiating objects.

In the aftermath of a Mass Casualty Incident (MCI) many patients require lifesaving treatments and surgeries. Due to the sudden surge in demand, the resources of the hospital are overwhelmed, making proper planning and use of the available resources crucial in minimizing mortality and morbidity. To help with planning, patients are triaged into four levels based on the clinical assessment of the criticality of their conditions. The triage decisions are however subject to error and a patient may be under or overtriaged. Mistriages can be identified by performing imaging, e.g., a Computed Tomography (CT) scan, but imaging also takes non-negligible time and has limited capacity. We propose a queueing network model of patient flow during an MCI and use simulation experiments to quantify the value of identifying mistriaged patients. Our results demonstrate the value of performing imaging, but also point out to the importance of accounting for its limited capacity.

This paper presents a parallel algorithm for the path planning of multiple unmanned aerial vehicles (UAVs) in the context of a surveillance mission. The UAVs are tasked to visit a set of points of interest (POIs) dispersed in a 3D environment and the algorithm allocates the POIs to the UAVs and computes optimal paths in between the POIs. The algorithm following a four-step approach and relies on a single source shortest path (SSSP) algorithm to compute the optimal paths between the POIs and a genetic algorithm to assign the POIs to the UAVs and find the order in which the POIs are visited. The algorithm is parallelized on a graphics processing unit and a multicore CPU to reduce the computing time and to allow for in-flight planning. The proposed algorithm is able to calculate paths for 3 UAVs and 10 POIs in just 0.6 seconds which represents a speedup of 48x compared to a sequential implementation on CPU.

Kinesthetic teaching and Dynamic Movement Primitives (DMPs) enable fast and adaptable learning of robot tasks based on a human demonstration. A task encoded as a dynamic movement primitive can be reused with a different goal position, albeit with a resulting distortion in the approach trajectory with regards to the original task. While this is sufficient for some robotic applications, the accuracy requirements for assembly tasks in an industrial context, where tolerances are tight and workpieces are small, is much higher. In such a context it is also preferable to keep the number of demonstrations and of external sensors low. Our approach relies on a single demonstration and a single force-torque sensor at the robot tool. We make use of a Convolutional Neural Network (CNN) trained on the force-torque sensor data to segment the task into several movement primitives for the different phases: pickup - approach - insertion - retraction, allowing us to achieve better positional accuracy when generalizing the task primitives to new targets. To the best of our awareness, we are the first to utilize a CNN as a segmentation tool to improve the generalization performance of DMPs.

Charge Trap Transistor One Time Programmable Memories were introduced in 14nm as a secure non-volatile memory. This paper highlights some of the test features and complexities that were modeled in a Verilog behavioral model.

For the route planning of electric vehicles (EV) a greater emphasize is placed on minimizing the energy consumption. The model used to calculate the energy along the path is typically based on the mechanical model of the vehicle. However, to be more accurate, one should also consider the motor losses. In this paper, we propose an integrated motor optimization and route planning for EV based on the Particle Swarm Optimization (PSO) and the Bellman-Ford (BF) routing algorithm. The PSO is used to calculate optimized magnetic flux settings for an induction motor for various operating points. The calculated settings maintain the high efficiency of the motor throughout the trip, but are also used to accurately calculate the motor losses prior to planning the route. The BF algorithm is used to calculate optimized routes. A Pareto front of optimized routes that minimized energy and distance is produced and allows the user to select the preferred route. Both the PSO and the BF are implemented in CUDA on an embedded NVIDIA Jetson TX2 graphics processing unit (GPU) for maximum performance. The system is tested on road maps with up to 4.6 million edges and provides a speedup of 19.3x for the PSO and 10.1x for the BF.

The 2018 IROS Fan Robotic Challenge tasked participants with programming a robot to autonomously open and close a Spanish folding fan, highlighting the obstacles still associated with the dexterous manipulation of objects for robotic systems. Since high DoFs grippers are complex to coordinate and overkill for many industrial processes, our approach used an under-actuated parallel gripper with a 3D-printed adaptation to precisely grasp the fan in such a manner that gravity could be leveraged to act on the fan to produce an extrinsic, or external, dexterity. With our approach, we completed the challenge in 12.38 seconds, resulting in a top three finish. Furthermore, using a multi-modal tactile sensor, we analyzed the vibrations in the grasp during the manipulation and were able to distinguish the opening and closing of the fan from the motion of the robot with a 83% accuracy.

An 8Kx192b One Time Programmable Memory (OTPM) is designed and manufactured in a 14nm FinFET technology without process adders or additional masks. A Differential Current Sense Amplifier (DCSA) and dynamic adaptive 192b parallel programming results in a < 3 PPM bit error rate, and supports a 10 year 105°C retention specification.

DC machines have been in operation for over 100 years. Engineers and technicians understand the DC motor structure and components. They also have experience in rebuilding, testing, and making performance adjustments to the DC machines. Those that truly understand commutation, the commutation zone, and the characteristics of the carbon brush are, alas, few in number. Traditionally, the lack of a visible arc has been the defining quality for good commutation - hence the term “black”. The lack of proper understanding and relatively high-level physical science involved has led to the art of achieving this condition being nicknamed “Black Magic”. It is a reference to both the brush coloration, and commutation. The purpose of this paper is to define the commutating zone and present the critical components that create the zone. An explanation of specific brush properties and their contribution to successful commutation is included.

One of the most fundamental element of the Unmanned Aerial Vehicles (UAV) path planning is the assurance that the path planned avoids any terrain collision. The task to “evaluate how much a UAV path is in collision with the terrain” is particularly crucial for path planners that are optimizing random generated paths. This terrain collision evaluation task can be computationally demanding and its computation time depends on the environment representation used. This paper presents a Field Programmable Gate Arrays (FPGA) based design of a UAV terrain collision evaluator that evaluates in parallel segments of the path in real-time. Its worst case computation time has a fixed upper bounded, based on the size of the map. For a 500×500 grid map, this upper bound is 1.9 μs, which is a promising result for future UAV real-time path planners.

Traditional route planning algorithms, such as Dijkstra or A∗, are efficient in their execution due to low computational complexity. With the rise in popularity of electric vehicles, which have the capability of recharging their batteries while driving, energy costs to traverse road segments can now be negative values. Dijkstra or A∗ cannot process negative edge weights. This work presents the use of the Bellman-Ford algorithm to plan driving routes based on energy efficiency. To overcome the increased computational complexity and ensure reasonable processing speeds this work uses an embedded graphical processing unit and a parallel implementation of Bellman-Ford on an embedded GPU system. This method allows routes to be plotted with considerable reductions in energy requirements, while maintaining the performance of traditional route planning programs.

Today, collaborative robots are often taught new tasks through “teaching by demonstration” techniques rather than manual programming. This works well for many tasks; however, some tasks like precise tight-fitting insertions can be hard to recreate through exact position replays because they also involve forces and are highly affected by the robot's repeatability and the position of the object in the hand. As of yet there is no way to automatically detect when procedures to reduce position uncertainty should be used. In this paper, we present a new way to automatically detect insertion tasks during impedance control-based trajectory teaching. This is accomplished by recording the forces and torques applied by the operator and inputting these signals to a convolutional neural network. The convolutional neural network is used to extract important features of the spatio-temporal forces and torque signals for distinguishing insertion tasks. Eventually, this method could help robots understand the tasks they are taught at a higher level. They will not only be capable of a position-time replay of the task, but will also recognize the best strategy to apply in order to accomplish the task (in this case insertion). Our method was tested on data obtained from 886 experiments that were conducted on eight different in-hand objects. Results show that we can distinguish insertion tasks from pick-and-place tasks with an average accuracy of 82%.

As technology develops, manufacture process becomes more and more automated using robots. There is demand for high performance tactile sensor which can support robotic grippers in manipulation tasks especially for unstructured flexible objects. Despite the efforts that have been spent, the fabrication process of those functional sensor remains complicated due to their requirement of specialized materials and equipment. The proposed multimodal sensor overcomes the difficulty by enhancing the electrical and mechanical design therefore simplifying the manufacture steps. In this version, static and dynamic sensing are integrated in the same layer of capacitive sensor with direct written microstructured dielectric. This structure allows it to have large range of force sensing as well as the ability of detecting contact events such as slippage or losing of contact.

Grasping tasks have always been challenging for robots, despite recent innovations in vision-based algorithms and object-specific training. If robots are to match human abilities and learn to pick up never-before-seen objects, they must combine vision with tactile sensing. This paper present a novel way to improve robotic grasping: by using tactile sensors and an unsupervised feature-learning approach, a robot can find the common denominators behind successful and failed grasps, and use this knowledge to predict whether a grasp attempt will succeed or fail. This method is promising as it uses only high-level features from two tactile sensors to evaluate grasp quality, and works for the training set as well as for new objects. In total, using a total of 54 different objects, our system recognized grasp failure 83.70% of time.

Unlike commercial airlines that fly predetermined trajectories, military unmanned aerial vehicles (UAVs) operate in dynamic environments and must often adjust their itinerary based on the developing conditions during the mission. The path planner module is a key element of any autonomous UAV. It computes the optimal path from a start point to an end point. In this paper, we present a parallel genetic algorithm for UAV path planning using an embedded NVIDIA Jetson TX1 single-board computer. The path is built as a series of line segment connected by circular arcs to remove discontinuities and to account for the dynamics of fixed wing UAVs. It is optimized to minimize the average altitude avoiding detection by enemy radars and to minimize fuel consumption improving range. The software developed is tested on four different 3D terrains. By exploiting the parallel architecture of the Jetson TX1 GPU, the proposed path planner provides a speedup of 33x compared to a sequential execution on an ARM processor. It calculates quasi-optimal solutions in complex 3D environments in less than 4 seconds and requires only 10 Watts, making it an excellent solution for onboard path planning.

This paper explores the use of haptic and visual feedback during routine object-manipulation tasks. We conducted tests of slippage detection, contact detection, and grasp control, with the help of twelve human participants. During the tests, the subjects controlled the robotic arm of a UR5 Universal Robot that was mounted with a three-finger robotic gripper and equipped with tactile sensors. The participants used either their visual feedback, vibrotactile feedback, or pressure feedback to assess the movements of the robotic fingers as they attempted to complete each task. We analyzed their performance with each type of feedback in order to evaluate how haptic feedback may be used to reduce the need for visual attention, and thus improve the lives of upper-limb amputees.

Robotic operations that involve the displacement of objects generate different kinds of dynamic events. These may simply correspond to normal robot-related motion, or contact(s) with the object(s) during grasping, but they may also be potentially-problematic events like slippage. In this paper, we use sparse data from tactile sensors to detect slippage and discriminate object-gripper slip from object-world slip. The method we propose can also identify vibrations that correspond to other dynamic events automatically, even when those events are not related to slippage. The tactile data can then be classified, allowing the robot to react accordingly. To achieve this goal, we compute the power spectral density (PSD) of the tactile dynamic signal, and we apply transformations to the PSD that were inspired by the automatic speech recognition (ASR) field. The originality of this work comes from using a sparse representation of the transformed data to obtain sparse vectors containing a small set of high-level features. Those sparse vectors are then used as inputs to a simple linear support vector machine (SVM), that acts as a classifier and quickly estimates the event to which they correspond. Our method was tested on data obtained from 244 experiments that were conducted on 32 different everyday-objects. Results show that we can successfully discriminate most of the dynamic events we studied in this work. Moreover, by using this technique, we are able to detect slippage with an accuracy of 92.60% and to differentiate object-gripper slip from object-world slip with a success rate of 89.42%.

Power distribution networks are typically structured in a radial topology with extra tie switches to allow for a manual reconfiguration in case of unexpected failure or scheduled maintenance. With the implementation of the smart grid, it is now realistic to also consider the power demand fluctuation and have real-time reconfiguration of the network to always operate in the optimal topology, minimizing distribution losses. In this paper, we propose the use of a genetic algorithm to find the optimal configuration of the network. The algorithm uses a unique solution encoding based on branch weights and computes the minimum spanning tree to decode the candidate solutions. This novel encoding ensures that the radial topology of the network is maintained without the need for complex operators resulting in an efficient and powerful solver. Finally, the solver is tested on distribution networks ranging from 16 to 4400 buses. The quality of the final solutions is equal or better, the maximum network size considered is much larger and the execution time is significantly shorter than that of state-of-the-art methods.

This paper presents a robotic system that was used to study the restoration of touch sensitivity. This approach could improve the lives of people who suffer from having lost organs or upper-limbs. Here, a combination of tactile sensors, robotic fingers, and a haptic interface enabled us to undertake different types of experiments on human subjects. To this end, we have conducted two separate tests on eight human subjects in order to assess the effectiveness of the static and dynamic modalities in different detectable ranges of the skin sensitivity. As of now, overall results show the relative functionality of the proposed mechanism for further experimental research. Under a static condition we received better restitution feedback at the lowest and highest magnitude levels, compared to the two in-between levels. The lowest and highest magnitude levels, 2N and 8N respectively, had an overall success rate of 75%, while the two middle levels, 4N and 6N, had a success rate of 43%. Under a dynamic condition, results showed that the whole system could adequately convey texture information via vibrations, and that it allowed subjects to successfully differentiate textures 78.33% of the time.

Miniaturized, multi-material fibers represent a novel promising approach to integration of higher-level functionalities into textile mediums. In this work, we present novel "smart textile" fabrics featuring multi-material fiber antennas compatible with existing 2.4GHz wireless communication networks. A submillimeter leaky coaxial cable fiber antenna is shown to be fully compatible with a textile weaving process, and exhibits an isotropic radiation pattern in the H-plane and multiple radiation lobes in the E-plane, in accordance with theory and simulations. Keywords - wearable antennas, multi-material fibers.