RBE Capstone Presentation Showcase - Fall 2023

Friday, December 15, 2023
5:00 pm to 8:00 pm
Location
Floor/Room #
500

 

Screen B

Screen C

Screen D

Screen F

Screen H

Screen I

5:30 - 5:55

Uday Shankar

Venkatesh Mullur

Ankit Mittal

Peter Murray

Mayank Bansal

Abizer Patanwala

6:00 - 6:25

Nikunj Polasani

Sanya Gulati

Shreyas Chigururpati

Shivam Sharma

Prasana Vijay Natu

Aadesh Surendra Varude

6:30 - 6:55

Daniel Goto

Adri Rajaraman

Om Vinayak Gaikwad

Miheer Diwan & Siyuan Huang

Mihir Kulkarni

Deepak Harshal Nagle

7:00 - 7:25

Kyle Mitchell

Ronald Pfisterer

Thabsheer Jafer Machingal

 

 

 

             

Screen B

Presenter:  Uday Shankar - Rosendo (5:30 - 5:55)

Project Title:  Natural Command-Driven Robotic Manipulation using Large Language Models

Abstract:  This project, 'Natural Command-Driven Robotic Manipulation using Large Language Models,' innovates in the field of robotics by enabling robots to interpret and execute natural, human-generated commands. Utilizing advanced language processing technologies, specifically large language models, the system translates verbal or written instructions into structured commands for robotic manipulators. This approach enhances human-robot interaction by allowing more intuitive and accurate command interpretation, crucial for precise and safe robotic operations. The integration of this technology has wide-ranging applications, from industrial automation to assistive robotics, representing a significant advancement in making robotic systems more accessible and user-friendly, thereby bridging the gap between sophisticated robotic functions and everyday human communication.

 

Presenter:  Nikunj Polasani - Rosendo (6:00 - 6:25)

Project Title:  Path planning algorithms for mobile manipulators

Abstract: This capstone project focuses on path planning algorithms for robotic manipulators using NVIDIA Isaac Sim. The goal is to develop efficient and safe path planning solutions in realistic physics simulation and environments, which can be transferred to physical robots and real time applications.

 

Presenter:  Daniel Goto - Rosendo (6:30 - 6:55)

Project Title:  Real-time Energy consumption in Isaac Sim

Abstract:  Simulating energy efficiency in robot design allows rapid design iterations while taking into account battery life, heat generation, and operating costs. In this research, an extension was developed for Isaac Sim, a robotics simulation platform being developed by Nvidia. This extension provides a convenient way for users to input motor parameters, view real-time energy consumption, and save energy data from simulations. The extension can be easily loaded from the extension window and works with any articulated asset in Isaac Sim. This will help researchers achieve the goal of developing more energy efficient robots.

 

Presenter:  Kyle Mitchell - Rosendo (7:00 - 7:25)

Project Title:  Making Music with a Robotic Glass Harp

Abstract: Most people see robots purely for their ability to complete tasks. However, there is an often unconsidered area of robotics that highlights robots’ ability to be artistic. To help widen this area of robotics, a new robotic musical instrument was created. This paper describes the design, implementation, and purpose of an autonomous glass harp robot. Until now, there has not been a fully autonomous glass harp robot produced. Only systems that have mechanized glasses for rotation but rely on a human’s touch to play each glass have been created. The autonomous robot is able to interpret MIDI note data from Ableton Live and process it to play notes within an octave and a half of the chromatic scale (D#5 to G#6). In addition, the robot is able to accept human input via a MIDI keyboard so that humans are able to play the robot as an instrument, as well. The success of this robot lies in its ability to fill a gap within the area of musical robotics and demonstrates a robot’s ability to produce art. Future work can include (but is not limited to) making the discussed upgrades to the current robot, creating a wider range of playable notes to make the robot more versatile, creating a system to change the amount of water in each glass to add the ability for portamento notes (sliding a note from one pitch to another), or adding the singing wineglass robot to a band of other musical robots and composing pieces of music to demonstrate the ability for a group of robots to produce more complex art than just a standalone robot.

Screen C

Presenter:  Venkatesh Mullur - Calli (5:30 - 5:55)

Project Title:  Markerless Vision-Based Control in Occluded Environments

Abstract:  This study introduces a sophisticated method for synthesizing authentic robot configurations within occluded environments, employing a hybrid framework that seamlessly integrates an Attention U-net as a generator and a PatchNet as a Discriminator in the context of Wasserstein's Generative Adversarial Network with Gradient Penalty (WGAN-GP). The primary objective is to address the challenges posed by occlusions in robotic environments, where joints may be partially or fully obscured. The significance of considering occluded images lies in their common occurrences within real-world robotic applications, where environmental clutter and obstacles often lead to partial or complete occlusion of robotic joints. Addressing this challenge is crucial for tasks such as motion planning, control, and manipulation, where accurate visual information is vital for successful operation. The proposed methodology contributes to overcoming the limitations posed by occluded images, thereby improving the robustness and reliability of robotic manipulation systems.         

 

Presenter:  Sanya Gulati - Lewin (6:00 - 6:25) 

Project Title:  Hardware in the Loop Automation

Abstract:  The peripherals attached to Nike Station at Amazon Robotics are being tested manually. To save the time and effort of humans and to avoid human errors, we are developing hardware based tests for all peripherals that will be controlled by a cloud orchestrator.

 

Presenter:  Adri Rajaraman - Fichera (6:30 - 6:55)

Project Title:  Kinematic Verification for Continuum Robots via Optical Tracking

Abstract:  Concentric Tube Robots (CTRs) are miniaturized continuum robots that are typically studied for their potential applications in minimally invasive surgery. Kinematics for a CTR is unconventional and rely on using beam mechanics which introduce an intermediary set of robot parameters called arc parameters. The tubes themselves are often manufactured out of Nitinol, a super-elastic material that works well for this application. This research aims to validate the kinematics for CTRs that use Nylon tubes, which are cheaper and more easily sourced than Nitinol. Optical tracking using a stereo camera setup, from NDI, is used to track the robot tip position for the kinematic verification. Validation of kinematics for CTRs is important, so that can further research into developing differential and inverse kinematics for these robots is possible.

 

Presenter:  Ronald Pfisterer - Rosendo (7:00 - 7:25) 

Project Title:  Soft Robotic Orbital Debris Collector: A Novel Approach for Efficient Space Debris Management

Abstract:  The increasing presence of space debris in Low Earth Orbit (LEO) and the Federal Communications Commission’s (FCC) proactive stance on ensuring debris is placed in the proper orbit presents an urgent challenge for orbital missions, necessitating innovative removal strategies. This paper introduces a novel Soft Robotic Orbital Debris Collector, a design that leverages the capabilities of soft robotics, 3D printing, and foldable mechanisms. Instead of traditional joints to create this folding motion, soft 3D printed thermoplastic polyurethane (TPU) joints afford the flexibility to conform to various satellite geometries while maintaining resilience against collision forces and the G’s of rocket launches. Unlike traditional methods, this design envelopes and secures debris completely, irrespective of its shape or size, permitting safe and controlled repositioning or disposal by the satellite system. Drawing insights from foundational literature, the paper underscores the potential advantages and efficiency of this new approach in actively mitigating space debris. This innovative system's practical implementation, performance, and future potential are discussed, offering a fresh perspective on addressing the orbital debris challenge.

Screen D   

Presenter:  Ankit Mittal - Calli (5:30 - 5:55) 

Project Title:  Object Grasping through Shape controlled Variable Length Continuum Robot

Abstract:  In this project, we explore the practical advantages of configuring a continuum robot for adaptive object grasping by integrating a specially designed finger-type gripper into its structure. This innovative gripper utilizes cord actuation and combines soft robotics principles with silicon materials, showcasing its ability to effectively grasp objects of various geometries. Through compelling demonstrations, we highlight the robot's proficiency in manipulating a diverse range of objects, including those with different shapes and sizes, accomplished through the implementation of a clothoid-based visual servoing method. This research underscores the crucial role of configuration manipulation in continuum robots, offering adaptability and versatility in real-world applications such as medical procedures and household tasks, thus making a significant contribution to the advancement of robotics technology.

 

Presenter:  Shreyas Chigurupati - Calli (6:00 - 6:25) 

Project Title:  Depth Image Enhancement for Robotic Grasping: A GAN-Based Approach 

Abstract:  This project advances the application of Generative Adversarial Networks (GANs) for the enhancement of depth images in robotic grasping scenarios. Targeting the inherent limitations of depth sensors, such as noise and missing data, the proposed system utilizes a GAN model to refine the quality of depth images obtained from RealSense cameras. The enhanced depth images provide a more accurate representation of the environment, enabling more precise and dependable grasping by robotic manipulators. This approach is designed to significantly improve the robustness of grasp detection and execution in a variety of settings, including those with suboptimal lighting or complex object arrangements. This model offers a substantial improvement over traditional depth sensing methods.

 

Presenter:  Om Vinayak Gaikwad - Calli (6:30 - 6:55) 

Project Title:  Depth Refined: Innovative Depth Image Enhancement for Optimized Robotic Grasping 

Abstract:  In benchmarking analysis, it's crucial to not only compare algorithm performance but also to explore the variables influencing their effectiveness. Factors such as image quality, surrounding conditions, and the intricacies of the grasping algorithm itself play significant roles. The focus of this project specifically concentrates on image quality, hence, certain algorithms are constructed to process unrefined depth images to enhanced depth images.         
             
The methodology used in this project is to perform image processing techniques aimed at improving the resolution, contrast, and decreasing the noise levels of depth images currently captured from the Intel RealSense D435 camera. This includes the implementation of filtering techniques to reduce noise using Morphological Transformations, adjustment of dynamic range to enhance contrast using Contrast Limited Adaptive Histogram Equalization technique (CLAHE). Subsequently, the enhanced images are fed into existing grasping algorithms to assess the impact of the enhancements on their performance. Currently the enhancement is being tested on the GGCNN algorithm with a focus on its performance under varying lighting conditions, as outlined in the benchmarking methodology. This approach involves a detailed comparison of the algorithm's performance using enhanced images versus its performance with unrefined depth images. The aim is to quantify the improvements in grasp quality attributable to image enhancement, thereby offering a clearer understanding of the relationship between image quality and algorithm efficiency.

 

Presenter:  Thabsheer Jafer Machingal - Calli (7:00 - 7:25) 

Project Title:  Object Rearrangement for Robotic Waste Sorting 

Abstract:  The challenge of waste management demands innovative solutions, particularly in automating waste sorting processes. This project addresses this critical issue by exploring the efficacy of robotic object rearrangement on conveyor belts, a key step in automated waste sorting systems. The primary objective is to augment the discovery and segregation of waste items, facilitating more efficient sorting. To achieve this we introduce and evaluate three distinct methods: KMeans clustering, Principal Component Analysis (PCA), and Density Estimation. These techniques are employed to determine strategic start and end points for creating trajectories along which a robotic arm moves to rearrange objects on the conveyor belt.         
             
The crux of our research lies in the development and application of two novel metrics designed to measure the effectiveness of object discoverability and spatial distribution post-rearrangement. These metrics allow for quantifiable assessment of each method’s performance in spreading objects apart, thus enhancing the potential for precise waste identification and categorization. By implementing these methodologies, the project aims to increase the accuracy and efficiency of waste sorting, minimizing manual intervention and fostering a more sustainable approach to waste management.         
             
The expected outcome is a robust framework that not only optimizes waste sorting but also serves as a scalable model adaptable to various contexts within the realm of automated material handling and sorting. This research not only contributes to the field of waste management but also advances the capabilities of robotic automation in complex sorting tasks.

Screen F   

Presenter:  Peter Murray - Barton (5:30 - 5:55) 

Project Title:  Miren: an Attentive Musical Siren 

Abstract:  Miren is a robotic instrument that captures and transforms ambient sound into compelling musical phrases. Using a rotational air pump within a perforated stator, Miren produces a loud wail characteristic of 20th-century civil defense sirens. The instrument's pitch and volume are dynamically controlled with a brushless motor and airflow valve, allowing for continuous pitch variation and a wide variety of volume envelopes. Designed to be played in an open outdoor environment, Miren is sent control signals wirelessly. To generate music, Miren uses a Markov chain altered to account for rhythm matching and control signal generation to produce novel melodies that exploit the instrument's musical affordances.

 

Presenter:  Shivam Sharma - Farzan (6:00 - 6:25) 

Project Title:  SLAM for navigation in an orchard 

Abstract:  In this project, conducted from August to December 2023 at Worcester Polytechnic Institute, we explored the effectiveness of advanced SLAM (Simultaneous Localization and Mapping) techniques in complex agricultural environments, specifically orchards. The study primarily focused on comparing the performance of ORB SLAM 2 and ORB SLAM 3 algorithms when paired with an RGBD camera system.         
          
We implemented these algorithms using the Robot Operating System (ROS), Gazebo for simulation, and a Husky robot. The study also integrated Inertial Measurement Unit (IMU) data with ORB SLAM 3 to evaluate its impact on algorithmic performance. Our key objective was to determine which SLAM system more effectively manages the unique navigational challenges presented by orchard environments, characterized by variable lighting, uneven terrain, and repetitive visual patterns.         
             
This study not only contributes to the growing body of knowledge in robotic navigation within agricultural settings but also provides practical insights for the deployment of autonomous systems in similar environmental conditions. The integration of IMU data with ORB SLAM 3 emerged as a key factor in improving trajectory estimation and environmental mapping, promising to enhance the capabilities of robotics in precision agriculture.

 

Presenter:  Miheer Diwan & Siyuan Huang - Sanket (6:30 - 6:55) 

Project Title:  Pixel Pathfinders: Deep Learning based Quadrotor Navigation in Uncharted Terrains 

Abstract:  The agility of small quadrotors makes them ideal for search and rescue missions, enabling them to cover vast distances swiftly and navigate dangerous terrain. However, achieving agile flight in cluttered environments such as forests and disaster zones is challenging due to the lack of prior knowledge about the environment. Navigating such environments relying solely on onboard sensing and computation, further amplifies the complexity of this task.         
             
We propose a method that utilizes deep learning to predict motor control inputs directly from images. This allows the quadrotor to navigate an unknown environment without any prior information about the surroundings and without relying on any external inputs from a GPS or a motion capture system. The network is trained using simulated data, and to assess its viability, we conduct a comparative analysis with ground-truth trajectories derived from optimal control methods. This offers us valuable insights into the potential effectiveness and advantages of our approach, recognizing the uncertainties associated with the proposed method.

Screen H   

Presenter:  Mayank Bansal - Zhang, Z. (5:30 - 5:55) 

Project Title:  Modifying VRCNet: A point-cloud completion network 

Abstract:  In the realm of autonomous systems, a diverse array of sensors, including LiDAR, lasers, and RGB-D scanners, are employed to gather 3D point cloud data, a critical component for environmental perception. Addressing the complex challenge of point cloud completion, which involves restoring missing points in these partial point clouds, is a crucial aspect of 3D computer vision technology. This research focuses on the advanced implementation of VRCNet, a system that utilizes a variational approach for efficient point cloud completion. The study extends beyond the basic implementation of VRCNet, delving into significant enhancements to its architecture. These refinements include the integration of batch normalization, and the innovative use of Swish and Mish activation functions. Additionally, it explores the impact of alternative loss functions, such as AdamW and RAdam, on the network's performance. The research also conducts thorough experiments with various learning rates, providing insights into their influence on the training process and the overall effectiveness of the results. This comprehensive approach not only strengthens the foundational capabilities of VRCNet but also offers valuable contributions to the field of 3D computer vision.

 

Presenter:  Prasanna Vijay Natu - Zhang, Z. (6:00 - 6:25) 

Project Title:  Integrating Strong-Star Convex Constraints into FlowNet for Enhanced Optical Flow Estimation with Deep Learning 

Abstract:  This study focuses on advancing optical flow estimation in computer vision through deep learning. We examine models such as FlowNetC, FlowNetS, SpyNet, LiteFlowNet, etc., which are key in tracking object motion in sequential imagery. Our approach enriches the FlowNet framework by introducing a strong star-convex loss function, leveraging the synergy of geometric principles and deep learning to enhance accuracy and efficiency in optical flow estimation. Our initial results are promising, suggesting potential improvements in the learning process and overall accuracy. This Directed Research offers a comprehensive view of the evolving landscape of optical flow estimation, highlighting the seamless blend of deep learning techniques with geometric insights.

 

Presenter:  Mihir Kulkarni - Zhang, Z. (6:30 - 6:55) 

Project Title:  Comparative Analysis of Feature Matching Algorithms in Computer Vision Applications:  SuperGlue vs LoFTR 

Abstract:  Feature matching is a crucial aspect of computer vision and plays a pivotal role in numerous applications such as image stitching, 3D reconstruction, object tracking, and recognition. Developing robust and efficient feature-matching algorithms is essential to overcome challenges like changes in illumination, viewpoint, scale, and occlusion. This paper presents a comprehensive evaluation of two state-of-the-art algorithms, SuperGlue and LoFTR on the IMC 2022 dataset along with derived novel insights.

Screen I   

Presenter:  Abizer Patanwala - Xiao (5:30 - 5:55) 

Project Title:  An architecture for on the move mobile manipulation 

Abstract:  This project introduces a generalised architecture for on-the-move mobile manipulation in environments characterised by uncertainty. Mobile manipulators,  capable of interactive tasks in diverse settings, rely on efficient and rapid task execution. The proposed architecture employs a global planner, Hybrid-A*, for path planning of the mobile base, coupled with a Regulated Pure Pursuit Controller for precise base control. Arm trajectories are generated using quintic polynomials. A novel approach for determining the optimal pick-up point is presented, integrating factors such as collision avoidance, arm manipulability, and maximum distance coverage within the manipulation region into a comprehensive cost function solved using the Differential Evolution algorithm. The system's efficacy is demonstrated through validation in the Gazebo simulation environment, showcasing its potential for enhancing overall task execution time in real-world scenarios.

 

Presenter:  Aadesh Surendra Varude - Xiao (6:00 - 6:25) 

Project Title:  An Advanced User Interface for Enhancing Human-Robot Interaction for Digsafe Marking Robot 

Abstract:  Developed a user interface designed specifically for layman utility workers, aimed at facilitating effortless guidance of the Digsafe robot to execute its primary tasks. This interface prioritizes the establishment of an efficient communication channel between the robot and the human operator. Key features incorporated into the interface encompass camera feedback, speed control mode, a comprehensive birds-eye map view integrating GIS data for seamless target location selection, and real-time display of the robot's live pose. These functionalities collectively empower the operator to effectively oversee the marking operation performed by the robot. Additionally, the interface provides essential information such as the robot's battery status and sensor data, offering crucial insights into the robot's operational parameters. This user-centric interface endeavors to bridge the gap between non-domain workers and robotic technology, enabling a more accessible and intuitive means of interacting with the Digsafe robot.

 

Presenter:  Deepak Harshal Nagle - Xiao (6:30 - 6:55) 

Project Title:  Multi-Level Semantic SLAM 

Abstract:  Modified SLAM algorithm previously developed by us offers a way for robots to perform the crucial task of aircraft inspection to comprehend their surroundings when using RGB- D sensors. Though this algorithm is robust in the environments with repetitive features, the actual process is quite slow. In this work, we further incorporated a multi-robot setup and a path planner to make the aircraft inspection more efficient, faster, and scalable. Additionally, the multi-robot setup helps the regions of interest during inspection to be captured with a denser point cloud for better comprehension of such regions.         
              
Index Terms—SLAM, RGB-D, aircraft inspection, multi-robot system, path planner

 

 

Department(s):

Robotics Engineering
Contact Person
Sharon Kelting