MEng End of Year Showcase 2026

Our capstone project focuses on improving Magnetic Particle Imaging (MPI) by addressing limitations in nanoparticle tracer quality. Current tracer synthesis produces inconsistent particles that reduce imaging resolution, preventing clinical-grade performance. To solve this, we developed MagStick, a geometry-controlled magnetic purification device that selectively isolates high-performance nanoparticles, followed by a silica encapsulation process that stabilizes and optimizes their structure. Together, this forms an automated purification and coating pipeline that improves tracer consistency and achieves approximately a twofold increase in imaging resolution compared to existing methods.

Our system integrates robotics to automate mixing, filling, and handling, improving efficiency and reproducibility in EHD printing, where performance is highly sensitive to ink composition.

Computed Axial Lithography (CAL) is a novel volumetric additive manufacturing process that prints parts in a single step rather than layer-by-layer, enabling more complex geometries and faster print times than conventional 3D printing. In its most basic form, CAL uses a visible-light projector to selectively cure a rotating vial of photopolymer to build a part. OpenCAL is an open-source printer that broadens access beyond controlled labs for educational, research and hobbyist use. The system is designed for easy assembly with minimal tooling and includes a complementary post-processing solution, CentrifuCAL, to efficiently separate printed parts from excess resin. This project lowers barriers to entry and enables broader research, education, and future innovation in volumetric 3D printing.

The project applies cryoforging in liquid nitrogen to titanium alloys to enhance materials properties. We use Electron microscopy to examine nanotwin microstructure density, and tensile testing to measure strength and ductility. This approach provides a cost-effective solution for critical structural components in the aerospace, biomedical, and marine industries

This capstone project develops an automated design framework for microchannel cold plates using topology optimization to improve thermal management in high-power electronics. By leveraging chip heat maps as inputs, the system generates optimized cooling channel geometries that outperform traditional straight-channel designs. The approach combines a coupled CFD and thermal solver with adjoint sensitivity analysis and MMA optimization to efficiently handle large-scale design variables. Additionally, the project incorporates realistic constraints by optimizing under fixed flow rate conditions while balancing thermal performance and pressure drop. Overall, the solution enables faster, more effective cooling design for modern data centers and high-performance computing systems.

Modern AI models are now used in high-impact areas like healthcare and law, but their decisions can be hard to explain, making it difficult for people to trust and safely use them.

Automated code transpilation using Large Language Models (LLMs) offers an efficient, less error-prone alternative to tedious manual code rewriting for addressing memory safety and runtime inefficiencies. Our research presents a generic framework for automated, validated transpilation using LLMs, demonstrating its feasibility for industry adoption. We validate our approach through two applications: CToRust, which transpiles C code to Rust, and PandaX, which optimizes Jupyter Notebooks for GPU and CPU backends.

Our Capstone Project explored operational transparency, focusing on how the goodwill generated by this transparency depends on when work is and isn’t done. By examining parcel delivery data, our research showed that memory limitations make customers particularly sensitive to the end of service operations, leading them to leave higher ratings when activities happen close to the delivery time. The project contextualized these findings using psychology’s peak-end rule and emphasized that people value the certainty of knowing when a service will be completed. Applying these research insights, our team ultimately built a 24/7 AI agent to answer client questions, utilizing our findings on transparency to optimize customer interactions and improve service satisfaction.

As technology advances, processes that heavily utilize energy and require extensive cooling are constantly pushing the boundary of available computing solutions. Our team is utilizing machine learning models to simplify the design of cooling systems that would be used on supercomputers and data centers. The design cycle includes physical modelling of novel heat exchangers, assessing their performance via computational fluid dynamics simulations, training a machine learning algorithm, and 3D-printing an optimized heat exchanger.

Teaching AI is an AI-driven education assistance. With the course material as an input, it provides personalized assistance for every student’s question. In this project, we are cleaning up the format/structure of materials across different courses to improve overall performance.

Our capstone project develops a Chinese Classical Music Generation Agent, a structured AI system for producing culturally authentic music at scale. Unlike generic music models, our approach uses a multi-stage pipeline—from composition planning to structure generation and audio synthesis—to ensure controllability and consistency. The system is designed to address the limitations of one-shot generation, particularly in domains requiring strong cultural and structural constraints. We validate its commercial potential through market analysis, competitive positioning, and a scalable B2B/B2C revenue model.

Our Capstone project, *OpenPaths*, develops an intelligent transit planning system that generates personalized and accessibility-aware travel itineraries for urban users, especially elderly and wheelchair users. The system integrates large language models with routing tools and real-time data (e.g., transit, micromobility, and accessibility datasets) to provide coordinated, end-to-end trip planning. A key contribution is combining routing optimization with accessibility evaluation using a generalized travel cost function and gravity-based accessibility metrics. This allows us to quantify equity impacts by comparing accessibility between baseline and wheelchair-constrained scenarios. Ultimately, OpenPaths aims to improve both individual travel experience and system-level fairness in urban mobility.

We are building a novel energy management system that leverages agentic AI and reinforcement learning to autonomously share energy in microgrids more efficiently. In parallel, we are building up a business model canvas to explore the viability of launching this technology as a startup.

Our project addresses the growing thermal management crisis in data centers by investigating Triply Periodic Minimal Surface (TPMS) lattice structures. TPMS structures offer high surface area compared to traditional strut lattices, or straight fins used in today’s data centers. Our novel methodology for creating TPMS structures at low cost (compared to previous methods) includes (1) generation of geometry, (2) polymer scaffold construction, and (3) nickel electroplating. Our foundation of simulations and construction pipeline will enable future researchers to create TPMS lattices at low cost for deployment in high-performance servers within Berkeley and beyond.

Over 2 million acres of land were affected by wildfires in California in 2021 alone. Reforestation is slow and costly, sometimes even physically impossible. Collaborating with the Morphing Matter Lab, our team developed a drone-based deployment system for their biodegradable seed carrier. Our system allows for precise deployment of 50 seeds across multiple geographical environments with a modular design that enables quick restocking between flights.

Desalination is emerging as a promising solution to increasing water scarcity, but membrane fouling remains a critical operational challenge that drives higher energy consumption and costly maintenance cycles. Our project develops a data-driven framework to continuously assess fouling risk using routine plant operational data, reducing reliance on manual testing. The framework supports more proactive maintenance decisions, such as cleaning or replacement, based on the predicted system condition. By moving beyond rule-based strategies, the approach aims to minimize operating costs while improving system reliability and membrane lifetime.

This capstone is a campus digital twin project centered on integrating infrastructure information into a unified interactive 3D platform. It is intended to improve how people understand campus systems by making spatial relationships, asset data, and scenario-based insights easier to explore and communicate.

The project aims to quantify the impact of biofouling on hydrodynamics and structural integrity on Principle Power’s offshore wind turbine platform. Additionally, it will involve the development of a parametric software tool that allows platform designers to access site-specific fouling load data within seconds.

Desalination is emerging as a promising solution to increasing water scarcity, but membrane fouling remains a critical operational challenge that drives higher energy consumption and costly maintenance cycles. Our project develops a data-driven framework to continuously assess fouling risk using routine plant operational data, reducing reliance on manual testing. The framework supports more proactive maintenance decisions, such as cleaning or replacement, based on the predicted system condition. By moving beyond rule-based strategies, the approach aims to minimize operating costs while improving system reliability and membrane lifetime.

Calectra has developed a silicon carbide thermal storage material that converts electricity into high-temperature heat that can be used for large industrial processes, targeting a sector responsible for ~20% of global CO₂ emissions. Our project designs a 10 MWh pilot system scaled 100× from an existing prototype, using ANSYS and SolidWorks to simulate heat transfer, fluid dynamics, and mechanical feasibility across varying geometries and insulation configurations. The goal is to validate a constructible, cost-competitive system architecture that bridges the gap between prototype and commercial deployment in order to decarbonize industrial heat.

This project aims to evaluate methane-to-hydrogen pathways for CO₂-free hydrogen production and then focus on a microwave-plasma solution based on an Evenson cavity. First, we will survey and critically compare competing low-carbon or carbon-free methane conversion routes, emphasizing reaction mechanisms, energy efficiency, byproduct management especially solid carbon, and scalability. Next, we will design and model an Evenson-cavity microwave plasma reactor for methane decomposition to quantify how plasma operating conditions influence methane conversion and hydrogen yield by COSMOL. The core question is whether non-thermal plasma can activate methane at relatively low bulk temperatures, decouple energy input from reactor wall heating, and suppress CO₂ formation while enabling stable carbon handling. Ultimately, the work will clarify the design space for CO₂-free methane pyrolysis and advance a concrete reactor concept that could integrate with waste-methane and distributed hydrogen infrastructure.

Earthquakes threaten communities and critical infrastructure, yet current wave simulation tools remain too slow and resource-intensive for rapid hazard analysis and large scenario studies, and they struggle to fully leverage the growing scale of seismic sensor data.
NeurDE aims to make acoustic wave modeling faster without giving up the physical reliability needed for hazard assessment, scenario screening, and data-driven seismology.
We pair the Lattice Boltzmann Method with a learned component that improves the most failure-prone step, while keeping the core transport process grounded in physics, establishing a hybrid approach between numerical solvers and end-to-end learned models (PINNs, FNOs, DPOT, etc.).

Wave energy is one of the world’s most untapped resources that can be harnessed to capture near carbon free energy. However, the price of producing power is still more expensive than fossil fuel, and other renewables, due to the complexity of the marine environment. Our goal is to minimize the cost of harvesting energy from ocean waves. We are replacing rigid components with inflatables, and using existing technology- a point absorber. This wave energy converter (WEC) converts the kinetic energy of waves to drive a generator and create electricity.

We are scaling up the usage of engineered harmless bacteria to bind to Rare Earth Elements. These elements are currently primarly refined in China using biohazardous and unsafe solvants. This approach aims at developing a reusable, effective and selective extraction method.

For the UC Berkeley (M.Eng. EECS) x Siemens Capstone project, we address the problem where commercial buildings account for 30–40% of global electricity consumption, yet most operate on fixed energy schedules that ignore how grid carbon intensity fluctuates hour by hour. Our project, Carbon-Aware Energy Optimization Platform, addresses this gap by building a full-stack software platform that integrates Siemens Building X telemetry with real-time carbon-intensity and weather data to intelligently shift flexible building loads — like HVAC — toward lower-carbon windows. The optimization engine uses Mixed-Integer Linear Programming and Model Predictive Control to generate energy schedules that minimize operational emissions without requiring new physical infrastructure. A Python FastAPI backend drives the scheduling logic while a React dashboard lets operators visualize and compare baseline versus optimized cumulative CO₂ emissions in real time. The result is a functional prototype demonstrating how forecast-driven carbon intelligence can be practically embedded into commercial building operations.

Clinicians don’t have a universal method to predict coma recovery, leading to subjective predictions. Our team is developing machine learning models that objectify clinician decision-making and produce a prediction of a patient’s mortality, saving lives and resources. We are leveraging an EEG (brain signal) dataset of 1,000+ patients across multiple countries, combining models trained at scale on high performance computing clusters.

Utilizing light-based 3D printing techniques, our team sought to bring 3D printing capabilities to an X-ray-based medical imaging machine known as a Computed Tomography (CT) scanner. Our project consists of material research and testing, micrometer-scale manufacturing, and a neural network-based image reconstruction algorithm. Our research will pave the way for commercial in-house 3D printing for dentists and other medical professionals who utilize CT scanning technology.

Heart disease is the leading cause of death in the US and worldwide. Cardi-V utilizes the open-source software SimVascular to develop patient-specific cardiac models that simulate the dynamic behavior of individual cardiovascular systems. Using clinical imaging data, we constructed patient-specific vascular models and implemented physiologically realistic boundary conditions. We then conducted both 0D and 3D computational fluid dynamics (CFD) simulations to evaluate pressure and flow dynamics within targeted vessels. Our goal is to provide personalized and clinically actionable diagnostics that improve disease prediction and ultimately contribute to reducing cardiovascular mortality.

Our capstone project focuses on developing biomimetic 3D-bioprinted cartilage implants to improve treatment options for osteoarthritis patients. We design multilayer, macroporous scaffolds that mimic native cartilage architecture, combining β-TCP–doped and hydrogel layers to enhance osteointegration, mechanical performance, and tissue integration. These scaffolds are engineered to improve nutrient diffusion and support cell viability, enabling more durable cartilage repair. Ultimately, our goal is to create a minimally invasive, longer-lasting alternative that can delay or reduce the need for total joint replacement.

The project works with Starkey to explore the integration of physiological sensing into hearing aids by investigating whether in-ear microphones can detect cardiac signals and estimate heart rate in real time. It focuses on capturing acoustic signals from the ear canal and processing them into beats per minute using advanced signal processing techniques. The work demonstrates the technical feasibility and clinical value of continuous heart rate monitoring through hearing aids. By transforming hearing aids into health-monitoring devices, the project aims to enhance their functionality and support improved user adoption and health outcomes.

Brain tumor MRI diagnosis demands that radiologists synthesize hundreds of images across multiple 3D sequences and longitudinal studies, imposing substantial cognitive load and increasing the risk of human error during assessment of tumor progression or treatment response. Current VLM models lack clinically verified training data and are inefficient with the analysis of 3D imagery data. Our capstone evaluates vision-language models as assistive tools for dialogue generation, using an interactive interface that mirrors clinical reasoning. By benchmarking different state-of-the-art VLMs on real-world data, we aim to reduce workload and improve consistency while preserving radiologist judgment.

The primary treatment for pectus excavatum is the Nuss procedure which often results in complications due to incision size and contact with internal organs. Our team is developing ELEVEX, a medical device that provides a safer, simpler, and less invasive alternative to current surgical practices. Through collaboration with pediatric surgeons, patients, and other experts in the field, this approach will eliminate the risk of harming internal organs and avoid large incisions while enabling a simpler procedure that expands surgical adoption and treatment accessibility.

PTSD affects over 300 million people worldwide, however many cannot access regular treatment. We enable wearable, noninvasive, out-of-clinic neuromodulation therapy via a frontier beamforming device & pulse-echo adjustment system.

This project develops a microfluidic chip for simultaneous single-cell RNA and protein capture. Cells are settled into indexed PDMS microwells, lysed, and their contents partitioned in parallel — RNA onto magnetic beads and proteins into a polyacrylamide gel via electrophoresis. Spatial indexing preserves cell-to-biomolecule pairing throughout transfer and downstream sequencing and mass spectrometry analysis.

Driven by recent advances in machine learning, decreasing hardware costs, improving sensor quality and fidelity, and the growing demand for teleoperation systems in robotic research, surgical robotics, and high-precision manufacturing, this project addresses the limitations of current commercial solutions, which are often expensive, produce low-quality data, and are kinematically mismatched to the human hand. It aims to design a low-cost, PLA 3D-printed, human-centered exoskeleton interface featuring a dry-fit assembly and accurate kinematic alignment with the human hand. The system will capture high-fidelity finger motion data to enable intuitive teleoperation and support machine learning applications.

This project develops a wearable-compatible system that enables continuous, noninvasive metabolic health monitoring by analyzing sweat biomarkers. It combines electrochemical impedance spectroscopy (EIS) hardware with a robotic platform that generates controlled artificial sweat samples to produce high-quality training data. Instead of relying on fragile chemical coatings, the system uses machine learning to classify electrolyte compositions from impedance spectra. By integrating this approach with smartwatch-compatible components, the project demonstrates a scalable, reusable method for biochemical sensing. Overall, it bridges the gap between conventional wearable devices and real-time metabolic insight.

Spinal cord injury affects approximately 40 million people every year and 30-70% of upper limb assistive devices get rejected by the user within the first year due to complexity and lack of adaptability. We designed an AI feedback pipeline that takes patient profiles, including sizing and movement limitations to create the optimal personalized hand grasper design. Our solution incorporates simulated tasks in a MuJoCo environment and a three-phase LLM-guided optimization pipeline. Our goal is to accelerate the development of personalized and effective assistive technologies.

The project presents a hardware–software co-designed compressed ultrasound imaging system for 3D imaging. Generating unique pulses and echoes through a synthetic aperture, a miniaturized system using 8 PMUT sensors on a single PCB is a 100x reduction in the number of sensors from existing systems. Coupled with a computationally efficient forward model and GPU-accelerated reconstruction, the system accurately reconstructs a sparse 1 m3 scene at 1mm resolution in minutes rather than hours.

Oscillometric cuffs are intermittent and uncomfortable, while arterial catheterization, though accurate, is invasive and impractical for many settings. This project explores adapting PyrAmes Health’s existing continuous non-invasive blood pressure (cNIBP) technology for veterinary use, translating clinical and market needs into engineering requirements. The work lays a strong foundation for developing an integrated cNIBP device with broad applications in both veterinary and human medicine.

We are developing a piezoelectric platform that converts cyclic mechanical strain into localized electrical stimulation to mimic native electromechanical environment of the heart. Using PVDF films integrated into a deformable culture system, the device generates strain-induced electrical cues that directly stimulate cardiomyocytes. This system allows us to research how combined mechanical and electrical signals regulate cardiomyocyte maturation, function, and regeneration.

Over 100,000 adolescents in the U.S. are diagnosed with scoliosis annually, yet treatment effectiveness is often hindered by brace compliance, which averages only 58.8% of required wear time. To address this, our capstone project, BraceForward, integrates a low-profile, DIY piezoresistive pressure sensor system to implement into a scoliosis back-brace. This system utilizes an Arduino-based data acquisition setup to provide real-time monitoring of applied pressure, ensuring forces remain within the therapeutic window and providing objective feedback on both fit and wear time in order to improve clinical outcomes.

We are optimizing the quantity and quality of DNA origami structures. We are doing this by using a robotic liquid handling machine and incorporating gel electrophoresis with minimal human interaction. We are optimizing the protocol for DNA origami by testing ratios of staples and scaffold (the main reagents used in protocol). Findings can be used for applications in drug delivery and diagnostics.

Understanding conditions below the ocean surface is critical for coastal science, climate research, and environmental management, but collecting this data is often expensive and limited to short research deployments. Consequently, important subsurface changes can go undetected. To address this gap, our team is developing a low-cost autonomous sailboat for multi-depth ocean monitoring. The system combines autonomous navigation and a winch that profiles the water column using a CTD sensor and sediment sampler. By providing repeated measurements at different depths, the platform aims to frequently capture changes in the water column and make ocean monitoring more accessible and scalable.

Our Capstone Project is developing a low-cost autonomous driving platform that allows the robot to navigate to the destination and proactively avoid obstacles along the way using limited sensor input. It is built on the RDK-X3 robot car and the pipeline combines global path planning, onboard perception, and reinforcement learning. A key motivation of the project is to reduce reliance on heavy sensing, constant server communication, and expensive computing. We trained the model in simulation and deployed it in a real-world setting to evaluate how well it performs under chaotic driving conditions.

This project addresses critical labor shortages and safety risks in the construction and manufacturing industries by developing a framework for humans and quadruped robots to carry heavy objects together. Our system, known as HQ-PCoT, allows the robot to intuitively “feel” human intent and direction using only its built-in internal sensors, eliminating the need for expensive or fragile external hardware. By utilizing a hierarchical AI control architecture, the robot can dynamically coordinate its legs and arm to maintain stability while following a human partner’s lead. We successfully validated this approach through real-world deployment, demonstrating that a robot dog can safely co-transport a shared payload in a dynamic environment.

The project focuses on developing a tool to extract Helium-3 and other isotopes from lunar regolith. It uses a rake-like mechanism combined with vibration to loosen and release these materials from the Moon’s surface. The system is designed and evaluated through CAD modeling, material trade studies, and simulations to ensure performance in harsh lunar environments. The overall goal is to create an efficient method for collecting valuable resources that are scarce on Earth.

Robotic policy training can be slow and tedious but with imitation learning the process can be sped up tremendously. Imitation learning models such as ACT have millions of parameters fewer than the massive billion parameter reinforcement learning models and Vision Language action models currently used industry-wide. We discuss our results of training ACT policies on consumer grade hardware, testing our results on the Standard Open 101 arm with and intel RealSense depth camera, and what it means for the future of robotic policy training.

Healthcare facilities face high labor turnover rates (16.4% overall, 40%+ at long-term care centers) and costs from supply chain inefficiencies ($25B total annual loss in the US). Outdated internal logistics models are a systemic problem that forces clinical staff to spend over a third of their time on non-care-related tasks. We’re creating a mobile robotic platform that fully decentralizes the medical inventory, eliminating the need for nurses to retrieve items themselves. With medical-safe hardware design, advanced sensing and navigation, and multi-agent optimization via an integrated software platform, we’re approaching this problem from the perspective of what will actually help clinical staff rather than replacing them.

We are creating the chassis and navigation system for a larger project focused on reinventing healthcare logistics through autonomous robots. We manage the motion of the system, determining the dynamic model, and the autonomous navigation. For the navigation, we compare different methods, such as state-estimation based control, reinforcement learning control, and model predictive control, to determine which methods best work for a healthcare environment. We closely factor in the safety and privacy needs of our primary customer, inpatient facilities. We work closely with capstone group 6 to manage the business needs of the larger project, handling customer outreach, financials, and investor contacts.

Our project is about designing a machine-learning-based routing model to optimize communication links within a heterogeneous satellite network. It is designed to accomodate variables such as delay, signal quality, and available compute on the satellite nodes. To validate this model, we constructed a Hardware-in-the-Loop testbed using robotic vehicles to emulate real-world conditions the model might face.

Despite many technological advances in the field of robotics, there is still a need for the creation of linkage driven (non-tendon) robotic hands that reliably grasp a wide variety of objects. Our team seeks to address this problem by developing a flexible and dexterous robotic hand that uses motor-driven underactuated ulnar fingers to simplify grasping. By combining this linkage based under actuation with tactile sensors, we aim to demonstrate a practical approach to robotic grasping that can be applied across the industry.

State-of-the-art snake inspired robots are built for locomotion on dry, hard ground, with limited applications for traveling on granular media such as sand and soil. The UC Berkeley Embodied Dexterity Group needs a working model for researching undulating motion in sand. Our team built a multi-linkage limbless robot with a swappable head piece. The robot moves sinusoidally, with a waterproof “skin” that mimics the frictional properties of a snake’s ventral scales to propel it forwards in granular media.

Our team is using reinforcement learning to train a motion generation policy that controls a G1 humanoid robot. This policy allows a user to prompt the robot to walk or run mimicking a human reference by joystick command. This method creates a real-time, lightweight control policy that bypasses the need for torque level robotic control.

UC Berkeley’s Embodied Dexterity Group (EDG) has been developing soft robotic technology that uses sound frequencies to sense force. We have blended EDG research and art into an interactive, multi-sensory exhibit for the Exploratorium, a museum in San Francisco. Our nature-inspired exhibit includes air flow control, air compressor whistles, molded mushrooms, electronics, and robotic flowers. Our goal is to educate and inspire the public about current advancements in robotics.

We created a vision-based pipeline to generate machine-readable work-zone maps from monocular video sources (like dashcams). The system integrates traffic-cone detection, SLAM trajectory estimation, landmark anchoring, and geographic alignment, enabling scalable, cost-effective, and crowd-sourced work-zone mapping.

Satellites collect huge amounts of data every day, but slow and limited connections back to Earth make it hard to use that information in time. Working with our industry partner Satlyt, we’re building a system that lets satellites think for themselves using a small AI model running directly onboard. It quickly sorts through the data, flags anything unusual, and decides what’s most important to send back to the ground. The result: faster decisions in space and only the data that really matters making the trip home.

Flapping-wing drones serve as an excellent benchmark for the growing trend of increased complexity in aerospace systems, with their chaotic aerodynamics making them extremely hard to model and control. To address this, we used behavior cloning with real flight data to enable autonomous flight without the use of a physics model. Through combining a high-level motion controller with a neural policy that maps onboard sensor data directly to wing commands, we achieved robust repeatable flight on a platform where conventional modeling is often too costly or inaccurate.

Modern autonomous driving systems depend on real-time execution of perception, planning, and control algorithms on integrated onboard computing systems. Open-source autonomous driving simulators, such as CARLA, model the vehicle and environment in high fidelity, but abstract away how long it takes to process data, ignoring processor-level computation latency that can significantly affect performance.

The goal of our capstone project is to extend SHARC, a processor hardware and control algorithm simulator, and integrate it with CARLA. This builds a hardware-aware driving simulation framework, enabling us to systematically study how hardware constraints and computational delays impact autonomous decision-making in safety-critical scenarios, where the smallest of delays can lead to unsafe actions.

We are developing a real-time Model Predictive Control (MPC) system for autonomous racing in the Indy Autonomous Challenge. From offline code generation through on-car solving, the entire pipeline is designed to meet a hard deadline per control cycle. Validation spans unit tests, software-in-the-loop simulation on dSPACE, and physical track testing at Putnam Park.

Our capstone project focuses on improving Magnetic Particle Imaging (MPI) by addressing limitations in nanoparticle tracer quality. Current tracer synthesis produces inconsistent particles that reduce imaging resolution, preventing clinical-grade performance. To solve this, we developed MagStick, a geometry-controlled magnetic purification device that selectively isolates high-performance nanoparticles, followed by a silica encapsulation process that stabilizes and optimizes their structure. Together, this forms an automated purification and coating pipeline that improves tracer consistency and achieves approximately a twofold increase in imaging resolution compared to existing methods.

Our system integrates robotics to automate mixing, filling, and handling, improving efficiency and reproducibility in EHD printing, where performance is highly sensitive to ink composition.

This project presents MoFL, a unified pipeline that enables humanoid robots to learn tasks directly from natural language instructions without real-world demonstrations. By leveraging generative video models such as Veo 3.1 and Seedance 2.0, the system generates synthetic “imagined” demonstrations that serve as scalable training data. These videos are converted into executable robot behavior through 3D scene reconstruction, motion retargeting, and policy learning, enabling end-to-end text-to-motion generation. The system produces physically plausible whole-body control in simulation across a range of tasks and verified through real-world deployment. This work highlights the potential of generative models as scalable data sources and planning priors for robotics.