Profile Photo

Rosh Ho

Software Engineer/Roboticist

Columbia University

I'm a recent graduate student working on artificial intelligence, embedded systems and HCI. My current research focus is applying the latest works from academia onto assistive rehabilitative orthotics.

Email: r.ho [at] columbia.edu

Projects

Variations of hardware braces for stroke rehab

ReGlove: A Soft Pneumatic Glove for Activities of Daily Living Assistance via Wrist-Mounted Vision

ReGlove is a low-cost vision-guided pneumatic glove that enables assistive orthoses using wrist-mounted cameras and edge computing, achieving 96.73% grasp accuracy and 82.71% success on object manipulation tasks—all for under $250 with commercial components.

Diagram of JIF framework for robot learning from human demonstrations

Train Robots in a JIF: Joint Inverse and Forward Dynamics with Human and Robot Demonstrations

A dynamics-driven pre-training method that learns manipulation-relevant representations from multi-modal human demonstrations (vision and touch) without action labels. By jointly learning inverse and forward dynamics and fine-tuning with few robot demonstrations, it significantly improves data efficiency, generalization, and robustness in contact-rich manipulation tasks compared to prior visual pre-training methods

Robotic hand hovering above object with grasp identified

Achieving Dexterous Manipulation via Computer Vision for Stroke Patients

Improves stroke rehabilitation by enabling prosthetic and exoskeleton hands to automatically select appropriate grasps using vision. Combines RGB images with depth maps to recognize five daily-use grasp types, with results showing depth cues compensate for limitations in muscle signals.

Robot dog schematics

Designing & Building Quadruped Robotic Systems

Built a quadruped robot from scratch, including mechanical design, electronics assembly, and software integration purely with laser-cutting to reduce build time and optimized for lightweight structure.

Image of data collection setup with depth camera headmounted and glove for data collection

Sensor Array Development for Cross-Embodiment Pre-training

Developed a multi-modal sensor array for collecting human demonstration data, including RGB/depth cameras (Intel RealSense D435 and D405) mounted on the head and wrist, plus tactile sensing gloves. The ROS2-based system captures visual streams at 35Hz and converts depth data to point clouds for cross-embodiment pre-training.

Camera identifying insertion point of modular solar panel for insallation

Lightweight Vision Model for Solar Pole Identification

Developed a lightweight computer vision pipeline using YOLO to detect solar installation pole tips from RGB images with high precision. The model demonstrates strong localization accuracy suitable for fast, on-device real-world applications.

Library Capacity Website Case Study

Real-Time Library Occupancy Management System

Built 'CULib Study Spot,' a real-time occupancy tracking app for Columbia University Libraries that helps students locate available study spaces efficiently. The platform features live seat availability, quiet zone indicators, and floor suggestions, while providing librarians with an independent tool for managing capacity data.

Miniature man sitting on coins

Political Marketing NLP Analysis

A Columbia Business School political marketing analysis project investigating how candidates signal bipartisanship and elite credentials supervised by Professor Mohamed Hussein. Features LLM-based tweet classification and campaign website scraping to understand impacts on voter perception.