Non-Intrusive Load Monitoring (NILM) is a technique that analyzes total household electricity consumption from a single smart meter to identify, in real-time, the energy usage of individual appliances. Using machine learning, it disaggregates electrical signals to estimate appliance energy use, acting as a cost-effective alternative to submetering. 

A way to reduce bandwidth and improve robustness for Industry 4.0 predictive maintenance, focusing on anomalous sound detection (ASD) in machine operating audio. Instead of transmitting raw waveforms, the architecture uses a semantic encoder at the edge to extract and send only semantically relevant representations of machine state, while a receiver-side semantic decoder performs the heavier reconstruction/inference needed for anomaly detection. I intend to evaluate the system on the DCASE 2020 Task 2 benchmark and compared against two conventional bit-level pipelines (OPUS for audio and JPEG) under simulated noisy links using a BSC channel model with calibrated error at different SNRs. Results are analyzed using AUC/pAUC metrics, highlighting improved communication efficiency and robustness while maintaining effective anomaly-detection performance. 

Keywords:

The research interests include linear control, stochastic control, cyber-physical security, and applications of information-theoretic control to industrial systems. Current research includes investigating the limitations of feedback control from information-theoretic perspectives, semantic control, and the relationship between reduced-model closure and learned policies in reinforcement learning. 

Keywords:

This research presents a CPU-FPGA co-design to resolve computational bottlenecks in Deep Reinforcement Learning (DRL) control for mecanum Autonomous Mobile Robots (AMRs). Traditional embedded CPUs struggle to process high-dimensional vision data and execute DRL policies simultaneously without incurring severe control latency. Our architecture leverages the Xilinx Kria KR260 SoC to partition these workloads. The ARM CPU runs a camera-assisted Extended Kalman Filter (EKF) that fuses encoder, IMU, and LiDAR data for robust state estimation. Concurrently, the FPGA's Deep Learning Processor Unit (DPU) accelerates CNN visual feature extraction and the quantized DRL policy inference. Evaluated in ROS 2 simulations, this hardware-software partitioning minimizes inference latency, ensuring deterministic real-time responsiveness and superior tracking accuracy compared to conventional CPU-only pipelines.

Keywords: Mecanum Wheels, AMR, Fushion Sensor, Reinforcement Learning, EKF, FPGA Acceleration, KR260

The Self-Evolving AMR Framework (SEA-Framework) aims to bridge the critical "sim-to-real gap" in autonomous robotics by creating a continuous, closed-loop evolution system. Leveraging the high-fidelity physics of NVIDIA Isaac Sim and the massive parallel training capabilities of NVIDIA Isaac Lab, the framework enables robots to autonomously detect environmental novelties and trigger on-the-fly re-simulation.

This approach moves beyond the traditional "train-then-deploy" paradigm, allowing AMRs to synthesize new expert policies dynamically and adapt to unseen terrains or complex edge-cases with zero-shot transfer reliability. The project focuses on scaling robotic intelligence through a persistent knowledge base that evolves alongside the robot's real-world experiences.

Keywords: AMR, Lifelong Learning, Deep Reinforcement Learning, Sim-to-Real Transfer, Nvidia Omniverse, Digital Twin

The proposed system consists of a robotic manipulation framework that performs grid-based tactile exploration using a compliant end effector. An ABB IRB 1090 industrial manipulator, equipped with a spring-based gripper, probes the target surface while a vision system monitors spring deformation to infer contact conditions. The resulting binary contact map is processed externally to estimate connector port locations, which are then used to guide insertion. Wire terminals are manually loaded into the gripper prior to the insertion stage; all subsequent mapping, localization, and insertion operations are performed autonomously. The hardware stack consists of an ABB IRB 1090 manipulator with RAPID-based controller, a custom compliant gripper with probing tip, a Cognex 8200 industrial camera, and a Raspberry Pi 3B embedded controller running Python for data processing.

Keywords:

Picking up a DuPont wire from a fixture using an ABB 1090 manipulator and inserting it into a connector socket may sound easliy, but it presents several technical challenges, such as wire deformation, incorrect insertion angles, excessive speeds, and insufficient precision. The target of this project is to raise the success rate of automated wire insertion by integrating the manipulator with peripheral device, including pneumatic grippers, cameras, fixtures, and custom 3D-printed components.

Keywords: insertion task、automation、ABB 1090、fixture、camera、gripper、3D printer

A feedforward vibration suppression approach for PPP Cartesian robots used in Automated Optical Inspection (AOI) of printed circuit boards, targeting the residual oscillations that force the camera head to settle before each image acquisition. Instead of waiting for passive damping, a Zero Vibration and Derivative (ZVD) input shaper convolves the motion command with a calibrated impulse sequence to cancel the dominant belt-drive flexible mode before it develops. The shaper is tuned from accelerometer ring-down measurements across nine workspace configurations, revealing a position-dependent natural frequency that shifts with both Z-axis extension and Y-axis position. A per-configuration lookup table is deployed on an ESP32 microcontroller and validated experimentally across seven configurations. Results show RMS residual vibration reductions of 17–71% (mean 43.7%), a maximum peak reduction of 79.1%, and a 47% settling time improvement at the highest-amplitude configuration. A simulation robustness sweep identifies a frequency ratio window of 0.7–1.4 within which a single shaper design provides consistent suppression, offering a practical tuning guideline for reconfigurable workspace systems. 

Keywords:

Capstan drives offer near-zero backlash and smooth torque transmission, making them attractive alternatives to conventional gear reducers in precision robotic applications. However, rope slip, elastic deformation, and load-dependent behavior raise reliability concerns. This work evaluates a 3D-printed (PETG) capstan transmission and equivalent gear system driven by a Dynamixel PM42 actuator through four structured tests: baseline characterization, load response, creep, and accelerated endurance. Transmission error, motor current, and rope condition are monitored under a three-tier FMEA failure framework (IEC 60812). Weibull statistical analysis extracts characteristic life and failure mode parameters to provide a quantitative basis for industrial feasibility assessment.

Keywords: