Research Projects Completed by Students during Summer Semester 2018
A Generic Deep Neural Network Framework for Object Detection on Miniature Mobile Robots
Ivy Nuo Chen, Temple University
Faculty Advisor: Prof. Fuman Zhang, Electrical & Computer Engineering
Neural networks have been demonstrated to be an effective means for object detection. However, neural networks require a great amount of computational power. Therefore, neural networks are typically deployed on devices like graphics cards, which consume large amounts of energy. In addition, the equipment necessary to deploy a neural network in the field can be substantial and cumbersome, which hinders its ability to be utilized on mobile robotics platforms, especially smaller scale robots. In this study we propose a solution that enables power-efficient neural network computation with a discrete neural network accelerator. The methodology involves collecting an image dataset, labeling the data, training the labeled images, and implementing the trained neural network on the embedded on board computer of the robot with acceleration hardware. The performance and power efficiency of this hardware-accelerated neural network solution was then compared with conventional approaches. The neural network solution was also implemented on an aquaculture inspection robot to verify its viability in the field. Results show that this generic neural network framework is ideal for object detection on miniature robots, as it is lightweight and has relatively conservative energy consumption.
Exploring Techniques for Detecting Intent Recognition for Powered Knee and Ankle Prosthesis
Divya Chowbey, Duke University
Faculty Advisor: Prof. Aaron Young, Mechanical Engineering
Lower limb powered prostheses, in contrast to currently available passive devices, have shown potential to provide transfemoral amputees positive mechanical work for different ambulation modes. In improving the functionality of our powered knee and ankle prosthesis, we aim to detect user intent. More specifically, our objective is to extend the ambulation mode capabilities of transfemoral amputees by applying machine learning. In this study, the methods used to develop the ambulation mode estimator for our prosthesis will be outlined and compared to approaches previous authors have proposed. To collect training data sets, transfemoral amputees (n=2) traversed a terrain park circuit consisting of level walking, ramp ascent, and ramp descent. Data was acquired from mechanical sensors on the prosthesis which included three inertial measurement units (IMUs), one 6-axis load cell, and two joint encoders. In creating a solution that could be implemented in real-time and produce high accuracy, linear discriminant analysis (LDA) and support vector machines (SVM) were explored. The value added by sensors was analyzed: the single best sensor/classifier combination was foot IMU and LDA, with reductions in steady state error to 5.8% and transition error to 16.1%. The impact of creating an accurate ambulation mode predictor will enable the powered prosthesis to restore natural gait biomechanics to the user.
Bob bots: Studying Jammed Systems with a Smart Active Particle Swarm
Sebastian Echeandia, University of Notre Dame
Faculty Advisor: Prof. Daniel Goldman, Physics
The study of jammed systems in granular media has been traditionally studied by the application of stresses to understand the reaction of the system in response to the application of forces. However, a novel approach suggests that the implementation of smart, active particles in the study of an armed system can provide new insights about efficient unjamming of glassy states. Based on this, the research objective of this paper is to design and manufacture a robotic swarm for the study of jamming and unjamming of a system composed of smart, active particles. This paper describes the design and manufacturing process of the robotic unit for the aforementioned swarm called Bob bot. After the construction of the proposed design, it was found that the robot exhibited all desired behavior, and its manufacturing time was reasonable for it to be mass produced. Therefore, a swarm made out of Bob bots is feasible and suitable for experimentation in the study of unjamming of systems and other relevant topics in complex rheology.
Robotic Realization of curved Spacetime
Alia Gilbert, Arizona State University- Polytechnic
Faculty Advisor: Prof. Daniel Goldman, Physics
Analog gravity is used as a tool to explore physical phenomenon in the field of general relativity. The behavior of a self-propelled robotic car was mapped on an elastic membrane with a center depression to achieve a correlate of the trajectory of orbits in curved spacetime. A differential-driven robotic car was optimized for symmetrical weight distribution to drive on the membrane with same performance in clockwise and counterclockwise motions. Self-propulsion allows for more control over variables and longer trajectories than passive objects. We explored possible sources of error from the symmetry of the car and the properties of the membrane. The results lead to a greater understanding of creating an analog for curved spacetime and exploration of conditions affected in variants of gravitational asymmetry.
Gait Synthesis for Biologically-inspired Feline Quadruped Robot
Ian Gonzalez-Afanador, University of Puerto Rico - Mayaguez
Faculty Advisor: Prof. Patricio Vela, Electrical and Computer Engineering
Legged animals are better suited to locomotion over rugged terrain than conventional wheeled or tracked vehicles. Robots inspired by this locomotion mechanism and associated strategies stand to be well-equipped for applications requiring high-mobility such as search and rescue, reconnaissance and scientific exploration. Generating coordinated actuator signals for stable legged locomotion on these platforms is a complex task. We study the kinematic motion of feline animals and extract key kinematic features characterizing their walking gait. These gait features then serve as constraints in a nonlinear optimization problem whose solution is a kinematic motion plan to accomplish forward displacement of the body frame. The problem is transcribed to a large-scale nonlinear programming (NLP) problem by the Matlab-based Optragen framework and is then passed to IPOPT, a numerical NLP solver. Two variants of a quasi-static quadrupedal walking gait were synthesized and implemented on a feline-inspired robotic platform.
Exploiting Bistability for High Holding Force Density Reflexive Gripping
Rianna Jitosho, MIT
Faculty Advisor: Anirban Mazumdar, Mechanical Engineering
Robotic grasping can enable mobile vehicles to physically interact with objects for delivery, repositioning, or landing. However, the requirements for grippers on mobile vehicles differ substantially from those used for conventional manipulations. Specifically, aerial grasping requires rapid activation, high force density, low power consumption, and minimal computation. In this work we present a biologically inspired robotic gripper designed specifically for mobile platforms. This design exploits a bistable shell to achieve “reflexive” activation based on contact with the environment. The mechanism can close its grasp within 0.15 s without any sensing or control. No electrical input power is required for grasping or holding load. The reflexive gripper utilizes a novel pneumatic design to open its grasp with low power, and the gripper can carry slung loads between 10-30 times its weight. This new mechanism is described in detail including the kinematics, the static capability, the control structure and the fabrication. A proof of concept prototype is designed, built, and tested. Experimental results are used to characterize performance and demonstrate the potential of this new approach.
Developing a Rehabilitation Therapy System with Socially Interactive Humanoid Robots
Joy Pinckard, Jacksonville State University
Faculty Advisor: Prof. Ayanna Howard, Electrical & Computer Engineering
Technology has become increasingly important in our day-to-day lives. However, it is not commonly used in activities that require human interactivity, such as in physical therapy. In this work, we present the design of an interactive system created to examine the influence of a socially interactive humanoid robot on the motivation of its patient in a physical therapy session targeting children with Cerebral Palsy. The system encourages repetitive exercises known to improve upper-extremity motor function which were provided by a licensed physical therapist. The humanoid NAO robot was programmed to act as a physical therapist during the activity by providing adaptive feedback. This feedback encourages the user as they touch two targets in sequence. In a previous study, we examined the influence of NAO with a virtual reality game (Super Pop VR) and found that the NAO robot therapist was capable of providing corrective feedback to participants yielding results equal to that of participants who interacted with a human providing the feedback. In this work, we aim to further explore the integration of humanoid robots into rehabilitative scenarios by exploring corrective feedback given to participants through a system designed and situated entirely in the real world. In future work, the system presented here will act as a method to further evaluate the NAO robot’s efficacy to act as a physical therapist in rehabilitative therapy scenarios.
Learning Material Recognition on a PR2 Through Self-supervised Manipulation with Multimodal Haptic Sensing
Katie Sosnowski, University of Arizona
Faculty Advisor: Prof. Charlie Kemp, Biomedical Engineering
Material recognition provides robots with the capability of inferring properties of everyday objects. It also enables more precise communication between humans and robots, as people can specify which type or material of object they would like the robot to interact with. In this paper, we present a method for large-scale self-supervised collection of haptic data with a mobile manipulator that interacts with everyday objects. In addition, we present a multimodal haptic sensor that can be easily attached to an existing robotic end effector and is capable of sensing vibration, temperature, force, and capacitance data. We demonstrate this multimodal sensing approach with a PR2 robot, which collected self-supervised data from both prehensile and nonprehensile interactions using bins filled with a variety of objects in both foam and plastic categories.
Learning Material Recognition on a Jaco Through Self-supervised Manipulation with Multimodal Haptic Sensing
Mallak Taleb, Wayne State University
Faculty Advisor: Prof. Sonia Chernova, School of Interactive Computing
Material recognition plays a vital role in our understanding of and interactions with the world daily. Having the capability to recognize materials informs us on what type of grip may be used to pick up an object. Material recognition can also inform dialogue between humans and robots, enabling the user to specify which type of object they would like the robot to interact with. We introduce a method for large-scale self-supervised data collection of haptic data of everyday objects with a robotic arm. We present a multimodal haptic sensor that can be easily attached to an existing robotic end effector and is capable to measuring haptic features such as capacitance, force, temperature, and vibration. We demonstrate how a Jaco arm can leverage these multimodal haptic sensors to learn the materials of hundreds of objects across 6 material categories, including ceramic, foam, glass, metal, paper, and plastics.