Jason Jabbour

Coming Soon.

Abstract

Vision-Language-Action (VLA) models have advanced robotic capabilities but remain challenging to deploy on resource-limited hardware. Pruning has enabled efficient compression of large language models (LLMs), yet it is largely understudied in robotics. Surprisingly, we observe that pruning VLA models leads to drastic degradation and increased safety violations. We introduce GLUESTICK, a post-pruning recovery method that restores much of the original model's functionality while retaining sparsity benefits. Our method performs a one-time interpolation between the dense and pruned models in weight-space to compute a corrective term. This correction is used during inference by each pruned layer to recover lost capabilities with minimal overhead. GLUESTICK requires no additional training, is agnostic to the pruning algorithm, and introduces a single hyperparameter that controls the tradeoff between efficiency and accuracy. Across diverse VLA architectures and tasks in manipulation and navigation, GLUESTICK achieves competitive memory efficiency while substantially recovering success rates and reducing safety violations. Please visit GLUESTICK.com for more information.

Abstract

We introduce RobotPerf, a community-driven benchmark of ROS 2 kernels designed to evaluate real-time robotics performance across a broad spectrum of hardware platforms. Our benchmarking suite includes ROS 2 kernels spanning the full robotics pipeline evaluated on real-time critical metrics for robotic systems. Our benchmarking framework offers ready-to-use instrumentation and is easily extensible to accommodate the evaluation of custom ROS 2 computational graphs. Our benchmark draws upon expertise of leading hardware acceleration vendors to present a standardized robotics benchmarking methodology. RobotPerf will continue to grow as an open-source and community driven benchmark, guiding research and innovation in hardware-accelerated robotic systems. Please visit RobotPerf.com for more information.

Abstract

Automating robust walking gaits for legged robots has been a long-standing challenge. Previous work has achieved robust locomotion gaits on sophisticated quadruped hardware platforms through the use of reinforcement learning and imitation learning. However, these approaches do not consider the strict constraints of ultra-low-cost robot platforms with limited computing resources, few sensors, and restricted actuation. These constrained robot platforms require special attention to successfully transfer skills learned in simulation to reality. As a step toward robust learning pipelines for these constrained robot platforms, we demonstrate how existing state-of-the-art imitation learning pipelines can be modified and augmented to support low-cost, limited hardware. By reducing our model’s observational space, leveraging TinyML to quantize our model, and adjusting the model outputs through post-processing, we are able to learn and deploy successful walking gaits on an 8-DoF, $299 (USD) toy quadruped robot that has reduced actuation and sensor feedback, as well as limited computing resources.

Accomplishments Summary

• Utilized artificial intelligence to overcome limitations of cheap hardware in a constrained quadruped robot
• Modeled a reinforcement learning problem of a constrained robot in simulation using OpenAI Gym and PyBullet
• Investigated methods for reducing the Sim2Real gap in robotic systems with limited sensing feedback
• Applied TinyML optimization techniques to reduce the size of a reinforcement learning model by 1000%
• Implemented a pipeline for deploying a Stable Baselines model on an embedded system using TFLite

Abstract

Securing network traffic on US Navy ships protects a ship from cyber-attacks which can have serious consequences. Cyber attacks can disrupt a ship's operations, cause damage to systems and equipment, and pose a threat to the safety of the crew. A reinforcement learning agent was trained to analyze network traffic data and identify anomalies that may indicate an attack is underway. By detecting attacks early on, the agent can alert peer systems to initiate countermeasures and mitigate the threat.

Accomplishments Summary

• Designed and built a prototype AI system for securing network traffic on US Navy ships
• Simulated cyber-attacks on US Navy ships navigating to waypoints using ROS, Gazebo, and OpenAI Gym
• Trained a reinforcement learning agent to detect attacks on a US Navy ship’s network traffic and trigger alerts
• Created a Flask-based dashboard for visualizing network traffic and detected attacks

Abstract

We apply optimal control theory to study the stable equilibrium shapes of a Möbius band that is modeled as an anisotropic elastic rod. Although previous work has focused on determining the equilibrium shapes of a Möbius band, little is known about the mechanics of how a circular rod deforms into a Möbius band as its ends are twisted. In this talk, we will show that a circular rod can experience a pitchfork bifurcation during this deformation. Using the necessary and sufficient conditions for optimality from optimal control theory, we numerically analyze how the critical twist angle at which these bifurcations occur depends on the stiffness parameters of the rod. Then, for a specific choice of stiffness parameters, we explore the structure of these bifurcations. Our findings show that although the elastic rod can bifurcate before being twisted into a Möbius band, all solution branches resulting from the bifurcation converge to the same Möbius band shape.

Accomplishments Summary

• Applied optimal control theory to study bifurcations in the construction of elastic Möbius bands using MATLAB
• Computed the critical twist angle of a Möbius band using numerical methods including the Runge-Kutta method
• Discovered the emergence of a supercritical pitchfork bifurcation at points of instability
• Discovered that all branches emerging from bifurcations occurring before 180° lead to the same Möbius shape

Abstract

Dementia caregivers often report increased anxiety and depression. In order to improve the interactions between in-home patients and caregivers, and reduce strain on caregivers, we build a monitoring, modeling, and interactive recommendation system for caregivers for in-home dementia patient care. The system includes monitoring for mood by speech, building classifiers that work in realistic home settings, and supporting an adaptive recommendation system to reduce stress of the caregiver.

Accomplishments Summary

• Applied voice-based machine learning techniques to improve Alzheimer’s patient-caregiver relationships
• Led the development of a recommender system to provide adaptive recommendations to Alzheimer caregivers
• Modeled the recommender system as a contextual multi-armed bandit reinforcement learning problem
• Developed the recommender system using Python, SQL, XML RPC, AWS EC2, and HTML injection methods
• Worked with healthcare professionals at OSU to deploy the RL driven system in real-world caregiver homes

Description: Coming soon!

Description: Collaboration between IARPA, Lightmatter, Harvard University, and Boston University on photonic computing to accelerate AI workloads in autonomous vehicles.

Business Wire article

Description

QueueClass.com is a platform that aims to improve the office hours experience for both students and teaching assistants. Many university courses have different methods for organizing office hours, such as creating their own websites, using Google forms, or relying on spreadsheets, which can be confusing and time-consuming for students and teaching assistants. QueueClass.com offers a central location for organizing office hours for all courses, streamlining the process and saving valuable time for both parties.

Accomplishments Summary

• Built QueueClass.com to facilitate a smoother office hours experience for students and teaching assistants
• Wireframed the preliminary UI design in Figma to enhance usability
• Developed using PHP, JavaScript, SQL, Ajax, HTML, CSS, Bootstrap and hosted on AWS EC2
• Implemented real time server-client communication using Node.js and Socket.io

Accomplishments Summary

• Applied deep learning techniques to classify images of murder hornets and bumblebees
• Built a Convolutional Neural Network (CNN) and compared its effectiveness to pre-trained CNN models
• Utilized LIME to interpret the distinguishing features detected in an image

Abstract

At the University of Virginia’s School of Engineering, Thornton Courtyard is a small area with vegetation and wildlife. There are five drains in this area that release stormwater directly into local watersheds, such as the Rivanna River and the Chesapeake Bay. The current drainage design causes these waterways to be impaired by the unfiltered stormwater. The goal of this project is to improve the drainage system at this specific site and demonstrate how it contributes to larger regional drainage concerns in Charlottesville. This will highlight the importance of addressing these issues to protect the environment and local waterways.

Accomplishments Summary

• Proposed a sustainable and resilient infrastructure solution for stormwater management
• Developed a re-design proposal for UVA’s engineering courtyard using AutoCAD Civil 3D to transition traditional stormwater practices to stormwater best management practices (BMPs)
• Surveyed site, conducted a water drop analysis, developed a demolition plan, integrated BMPs to the site re-design, created a site grading report, and produced a cut/fill report
• Reduced the peak runoff rate and the site runoff coefficient by 46% in simulation