Embedded Files
Legged Locomotion Systems
Design and development of legged robotic platforms exploring how mechanical design, kinematics, and actuation shape locomotion performance. This work includes a quadruped walking robot with independently actuated legs and hand-designed gait patterns, as well as a Theo Jansen–inspired linkage-based walker that achieves motion primarily through geometry rather than feedback control. Together, these systems emphasize a mechanics-first approach to locomotion, where morphology and linkage design play a central role in simplifying control.
Building on these systems, I employed CAD-driven mechanism design and kinematic analysis to define leg architectures, transmission layouts, and joint configurations for both actively controlled and passive locomotion platforms. Simulation and analytical reasoning were used to study link lengths, reachable workspace, and foot trajectories, informing actuator placement and structural layout prior to fabrication.
System integration involved motor selection, structural design, and electronics packaging, followed by iterative prototyping and experimental testing to evaluate gait stability, repeatability, and robustness across operating conditions. Through these projects, I explored how morphology, linkage geometry, and actuation choices can reduce control complexity and improve locomotion performance in legged robotic systems.
Mechanical Intelligence and actuation
Compliant Gripper with Mechanically Intelligent Structures
Design and development of a compliant robotic gripper that leverages structural compliance and bistability to achieve robust grasping with simplified actuation. The gripper explores how geometry, material distribution, and mechanical response can encode grasp behavior, enabling adaptability to object shape without complex sensing or control. This work emphasizes a mechanics-driven approach to manipulation, where structure plays an active role in function.
Building on this concept, the gripper was developed using CAD-driven parametric design to define compliant elements, joint geometries, and contact interfaces that govern deformation and force transmission during grasping. Nonlinear finite-element analysis (FEA) was used to study large deformations, local stress concentrations, and stability transitions in the compliant components, informing design choices prior to fabrication. Parametric studies explored the influence of geometry and material stiffness on grasp range, force distribution, and snap-through behavior.
Prototypes were fabricated using multimaterial additive manufacturing, enabling spatial variation in stiffness to localize compliance and control deformation pathways. Experimental testing focused on evaluating grasp repeatability, robustness to object geometry, and mechanical response under loading, allowing direct comparison with simulation predictions. Through this project, the gripper serves as a case study in how mechanical intelligence and structural design can reduce reliance on active control while improving adaptability in robotic manipulation.
Bistable and Programmable Metamaterial Structures
Design and investigation of mechanically programmable metamaterial structures that exhibit bistability, snap-through, and controlled deformation pathways under compression. This work explores how geometry, material distribution, and boundary conditions can be used to encode predictable surface topology changes and localized actuation without embedded electronics. The emphasis is on using structural mechanics as an actuation and control primitive in soft robotic and adaptive systems.
Building on this theme, I developed a range of 2D metamaterial architectures using CAD-driven parametric design, including living hinge arrays, seesaw bar structures, and elliptical-hole lattices with graded and offset geometries. These designs were intended to produce asymmetric deformation, alternating surface topologies, and snap-through behavior when subjected to compressive loading. Geometry was systematically varied to study wavelength selection, deformation localization, and the interaction between stiff inclusions and compliant matrices.
Extensive nonlinear finite-element analysis (Abaqus) was performed to capture large deformations, post-buckling response, and stability transitions, with simulations used to predict force–displacement behavior and deformation modes prior to fabrication. Structures were fabricated using multimaterial additive manufacturing, enabling spatial control of stiffness through material placement. Experimental testing under quasi-static compression was used to generate stress–strain curves, visualize evolving surface topologies, and compare measured behavior against simulation predictions. Collectively, this work demonstrates how bistability and geometry-encoded mechanics can be leveraged to create programmable mechanical responses relevant to soft robotics, adaptive surfaces, and mechanically intelligent actuation.
Programmable Porous Metamaterial Structures
Design and analysis of architected porous metamaterials whose global mechanical response is programmed through hole geometry, orientation, and spatial arrangement. This work demonstrates how relatively simple geometric variations can produce distinct deformation modes, including auxetic lateral contraction, conventional compression, and shear-dominated responses under uniaxial loading. The goal is to systematically characterize these behaviors and develop models capable of predicting and utilizing geometry-induced deformation patterns for mechanically intelligent actuation and adaptive structures.
Building on earlier experimental and simulation studies, this work focused on developing a computational design and analysis pipeline for two-dimensional elliptical-hole porous metamaterials. Using CAD-driven parametric geometry generation, families of panels were created with controlled variations in ellipse aspect ratio, orientation angle, stagger, alternation patterns, and spatial gradients. These geometric degrees of freedom were designed to induce specific deformation mechanisms, including auxetic responses in alternating configurations and shear-dominant deformation in slanted or graded layouts.
Extensive nonlinear finite-element simulations (Abaqus) were performed under plane-strain compression to capture large deformations, ligament buckling, and evolving deformation fields. The simulations revealed clear qualitative trends: alternating ellipse orientations produced lateral contraction, vertically aligned ellipses tended toward symmetric compression, and slanted configurations converted vertical loading into macroscopic shear-like deformation. These behaviors were consistent with earlier experimental observations and highlight the strong coupling between local geometry and global mechanical response.
To enable systematic exploration, an automated geometry → simulation → data-extraction pipeline was developed, allowing batch analysis of multiple architectures and extraction of stress–strain behavior, reaction forces, displacement fields, and effective Poisson ratios. While this work primarily establishes the computational and analytical framework, the longer-term objective is to use the resulting datasets to train predictive models capable of mapping geometry to mechanical response. Such models would enable the intentional design of porous structures that exhibit desired auxetic, compressive, or shear-driven actuation modes, advancing the use of geometry-encoded mechanics as a programmable design tool.
Mechanism Design and rehabilitation Robotics
Design and analysis of parallel robotic mechanisms that integrate geometry, compliance, and control to achieve safe and precise motion for human-interactive applications. This work focuses on a 3-PSS / 3-UPS parallel manipulator, exploring how mechanism architecture and passive mechanical properties influence workspace, force transmission, and interaction behavior. The emphasis is on using mechanical structure as an intelligence layer that complements control, rather than relying on control alone.
Building on this objective, a modified Delta-style parallel robot was developed using a 3-PSS (Prismatic–Spherical–Spherical) architecture, with prismatic actuators replacing revolute joints to improve adaptability, load capacity, and controllability. The mechanism was fully modeled using CAD-based design, followed by analytical derivation of inverse kinematics, forward kinematics, and rate Jacobians to characterize motion and constraint behavior. Workspace analysis and mobility calculations were used to identify feasible operating regions and singularity-free configurations.
Beyond kinematics, the system was extended to include compliance, friction, and gravity effects within a Newton–Euler dynamic formulation, allowing the mechanism to be analyzed as a coupled mechanical–control system. Impedance and compliance control frameworks were derived to study how passive stiffness in the structure interacts with active control, revealing trade-offs between stability, bandwidth, and achievable interaction forces. Simulation results demonstrated how mechanically embedded compliance alters system response, improves safety, and enables smoother human–robot interaction in rehabilitation-oriented tasks. Collectively, this work highlights how mechanism design, passive mechanics, and control modeling can be co-designed to produce mechanically intelligent robotic systems suitable for human-centered applications.
Bio Insipred Locomotion and Reinforcement Learning
Development of bio-inspired locomotion systems in which reinforcement learning (RL) discovers efficient gaits directly from physics-based simulation, without hand-designed trajectories. Inspired by earthworm peristalsis and inchworm looping, this work investigates whether simple mechanical primitives, coupled with minimal rewards, can give rise to emergent, coordinated locomotion. The emphasis is on discovery rather than imitation, using learning to uncover locomotion strategies encoded by mechanics and environment interaction.
The primary system models an earthworm-like robot as a chain of 11 point masses connected by spring–damper elements, simulated under gravity with directional (anisotropic) friction. Each action locally modifies the rest length of a single spring over a fixed time interval, enabling the emergence of peristaltic wave propagation through coordinated segment contraction and extension. The reward function was deliberately minimal, prioritizing forward center-of-mass displacement, with penalties for excessive energy use, instability, and nonphysical behaviors such as flipping.
Multiple learning paradigms were evaluated to understand the role of memory and coordination in locomotion. A Multi-Armed Bandit (MAB) baseline demonstrated limited improvement due to its lack of temporal state awareness. Q-learning introduced state dependence and enabled basic coordination, while Proximal Policy Optimization (PPO) produced smooth, biologically plausible gaits characterized by stable wave timing, higher forward velocity, and reduced energy expenditure. Comparative analysis showed clear differences in convergence behavior, gait smoothness, and robustness, highlighting the importance of policy-based deep RL for continuous control.
The framework was extended to an inchworm-inspired model incorporating dynamic anchoring, where the head or tail alternately locks to the ground during high-curvature looping motions. This required explicit modeling of contact physics, collision handling, and phase-dependent constraints, revealing how mechanical anchoring and RL policies must co-evolve to produce effective locomotion. Key challenges included tuning anisotropic friction parameters, preventing reward exploitation, and enforcing stability through carefully designed penalties. Overall, this work demonstrates how reinforcement learning can act as a tool for discovering locomotion strategies, revealing the tight coupling between mechanics, environment interaction, and learned control in bio-inspired robotic systems.
Vision-Guided Mobile Robot for Industrial Automation
Design and implementation of a vision-guided mobile robot integrating motion control, camera-based perception, and a web-based Human–Machine Interface (HMI). The system demonstrates core principles of industrial automation, including remote supervision, event-driven control, and perception-triggered actions using a minimal hardware platform.
This project involved building a mobile robot driven by two independently controlled motors and equipped with an onboard camera, designed to operate as a small-scale industrial automation system. A web-based HMI, developed using the Tornado framework, enabled remote control and monitoring of the robot from a computer, tablet, or smartphone over a network connection. The HMI handled command input, system feedback, and basic event reporting, reflecting common practices in industrial automation interfaces.
For perception, the robot performed camera-based QR code detection, identifying a specific assigned code while ignoring others in its environment. Upon detection, the robot executed a predefined event sequence: stopping motion, signaling detection via a timed visual indicator, and then autonomously returning to its starting position. This behavior simulated real-world automation tasks such as inspection stops, inventory identification, and safety-triggered halts.
The project emphasized a systems engineering approach, requiring integration across mechanical motion, software control, perception, networking, and user interaction. Design decisions were informed by considerations common in industrial settings, including reliability, deterministic behavior, and clear separation between sensing, decision-making, and actuation. The resulting system demonstrated how relatively simple hardware can support robust, perception-driven automation workflows when designed with a cohesive system architecture.
MLP vs KAN: Function Approximation and Interpretability
This project reproduces and evaluates Kolmogorov–Arnold Networks (KANs) against standard Multi-Layer Perceptrons (MLPs) on controlled function-fitting tasks. We implemented the KAN architecture—including spline-based edge functions, grid extension, sparsification, and pruning—and compared performance and behavior under fixed training budgets. The work reproduces several qualitative trends reported in the original KAN paper while highlighting practical optimization and training limitations.
We implemented multi-layer KAN models with B-spline–based edge functions and compared them against MLP baselines on toy function-fitting datasets and selected Feynman symbolic regression benchmarks. KANs were trained using a grid-extension strategy, starting from coarse spline resolution and progressively refining the grid while interpolating learned coefficients. Sparsification and pruning were applied to study architectural simplification after training.
All models were trained under fixed optimization budgets using Adam, with matched parameter scales where possible. Performance was evaluated using RMSE, and scaling behavior was examined as model capacity increased. In low-dimensional settings, KANs often achieved comparable accuracy to MLPs with fewer parameters, while in higher-dimensional or more complex functions, results were sensitive to optimizer choice and training budget. Pruning substantially reduced KAN model size but typically increased error, with partial recovery after limited fine-tuning.
Overall, this project provides a reproduction-focused assessment of KANs, emphasizing how optimization, training regime, and dataset structure affect both performance and interpretability, rather than presenting best-case results.
Product Design, Medical Devices & Patented Mechanisms
Experience in end-to-end product design for medical and robotic systems, spanning mechanism design, CAD, prototyping, and system architecture. This work includes patented medical device concepts developed at Becton Dickinson, as well as user-centered designs for high-volume drug delivery and vaccination workflows. The emphasis is on mechanical simplicity, safety, and manufacturability under real-world constraints.
At Becton Dickinson, I contributed to the design of novel medical device mechanisms that resulted in granted patents, focusing on mechanical architecture, CAD modeling, and design-for-manufacturing considerations within regulated healthcare environments.
I also worked on FlowSure, a modular on-body drug delivery system for large-volume infusions, where mechanical design centered on a roller-based pouch compression mechanism and a disposable–reusable system architecture to improve usability and sustainability. In parallel, I contributed to SMIP (Smart Multi-dose Injectable Pen), a system designed to improve efficiency and traceability in high-throughput vaccination settings, with emphasis on intuitive operation and dosing reliability.
Across these projects, I focused on translating user and clinical needs into robust mechanical systems, balancing performance, manufacturability, and user experience.
Exploratory Soft Actuation
Exploratory work on soft, sheet-like robotic actuation using Dielectric Elastomer Actuators (DEAs), investigating how material choice, electrode design, and geometry influence deformation and controllability. The project focused on understanding practical constraints and failure modes in DEA-driven systems rather than achieving a finalized robotic platform.
This project investigated DEA-based actuation as a mechanism for compact, compliant robots that deform through electro-mechanical coupling rather than rigid joints. Multiple elastomer substrates and electrode configurations were tested to study actuation response, breakdown behavior, and repeatability under high voltage.
In parallel, CAD-driven sheet geometries with stiffness contrasts and structural ribs were designed to bias deformation and induce bending when actuated. While sustained locomotion was not achieved, the work identified key limitations related to material durability, electrode reliability, and electric field concentration, informing future directions for integrating soft actuation with mechanically intelligent structures.
Foundational Engineering Projects
Early hands-on projects spanning aerial robotics, control systems, bio-inspired mechanisms, and experimental prototyping. These works focus on building intuition in mechanical design, dynamics, and system integration, and form the foundation for later research and robotics projects.
This collection includes team-based and individual engineering projects developed during earlier stages of my training. Projects range from multi-rotor aerial systems (ball drone, agricopter) and a self-balancing bicycle emphasizing control and dynamics, to bio-inspired mechanism design such as an ornithopter wing and linkage-based prototypes. While exploratory in nature, these projects involved full build–test–iterate cycles, including CAD, fabrication, electronics integration, and experimental validation, and helped develop a strong intuition for how mechanical systems behave in practice.
Page updated
Google Sites
Report abuse