My Work

Projects

From hardware robots to simulation frameworks to deployed software — a collection of things I've built and broken (and fixed).

Hardware 2025 – Present

Nursing Robot Perception Stack

End-to-end perception system for a bi-manual mobile nurse assistant — multi-camera calibration, depth fusion, object recognition, and shared autonomy on Jetson Nano.

<5% pose error ~60% less manual intervention
ROS2 PyTorch Jetson Nano Intel RealSense
View Project →
Hardware Nov 2024 – Present

6-DoF Pose Estimation for Medical Objects

Marker-less 6-DoF pose estimator using vision-only pipeline and multi-view fusion for bottles, syringes, and trays — no physical markers needed.

~30% less setup time ~2.0s/frame on Jetson
PyTorch TensorFlow RealSense D435 ROS2
View Project →
Hardware Dec 2024 – Feb 2025

CrazyFlie 2.0 Quadrotor Control

Cascaded PD and LQR controllers for a physical quadrotor with polynomial trajectory generation. Achieved sim-to-real transfer with under 7mm flight RMSE.

<3mm sim trajectory error <7mm real-world RMSE
Python MATLAB LQR PD Control
View Project →
Hardware Jun – Oct 2023

Gait Planning for Quadruped Robot

Inverse kinematics and adaptive gait controllers deployed on Raspberry Pi, improving stability on uneven terrain and dynamic balance under perturbations.

~30% better terrain stability 5% positional error
Python Raspberry Pi ROS2 PyBullet
View Project →
Hardware Oct – Dec 2024

Dynamic Control of OpenManipulator X

LSPB trajectory planning and PD joint control for a 6-DOF manipulator arm. Torque estimation validated within 5% of measured values using motor current profiles.

<5% path tracking error ~5% torque accuracy
ROS2 MATLAB LSPB PD Control
View Project →
Hardware Oct – Nov 2025

Cone Calorimeter HRR Computation

Full software pipeline to calibrate 14+ sensors and compute Heat Release Rate in real time using oxygen-consumption calorimetry with live monitoring dashboards.

>30% faster calibration 14+ sensors
Python NumPy SciPy Pandas
View Project →
Simulation Feb – Apr 2025

RL Control for UR10 Robot

DDPG and A3C implemented from scratch for pick-and-place manipulation on a simulated UR10. Includes imitation learning pipeline for peg-in-hole tasks.

~85% DDPG success rate ~55% IL baseline
PyTorch DDPG A3C Gym
View Project →
Simulation Mar – May 2025

Imitation Learning with ACT + Transformers

Action Chunking Transformer with VAE latent encoding trained on expert demonstrations. Outperforms behavior cloning by 30% on constrained peg placement.

+30% over BC baseline Generalizes to unseen configs
Transformers VAE PyTorch Imitation Learning
View Project →
Simulation Dec 2024 – Jan 2025

Visual-Inertial Odometry & Synthetic Dataset

10,000+ frame synthetic Vision+IMU dataset created in Blender. MSCKF filter + deep fusion network achieving <5% trajectory error, 20% better than vision-only.

<5% trajectory error +20% vs vision-only
MSCKF Blender PyTorch IMU Fusion
View Project →
Simulation Sep – Nov 2024

Gesture-Controlled Robot for Accessibility

Simulation and control stack enabling non-verbal gesture control of a robot in Gazebo. Customizable gesture mapping with 90%+ detection accuracy.

>90% gesture accuracy ~40% faster adaptation
ROS2 Gazebo MediaPipe OpenCV
View Project →
Simulation Jan – Feb 2025

Homography & Panoramic Stitching

Classical ANMS + RANSAC pipeline alongside HomographyNet trained on MSCOCO. Unified library for high-quality panorama stitching achieving ~1.5px RMSE.

~1.5px RMSE classical ~2px deep homography
OpenCV HomographyNet PyTorch RANSAC
View Project →
Simulation Feb – Mar 2025

Hybrid Boundary Detection & Classification

Two-phase pipeline fusing classical filters (DoG, Gabor, Canny) with CNN backbones. DenseNet achieved 92.5% CIFAR-10 accuracy using 30% fewer parameters.

92.5% CIFAR-10 accuracy 30% fewer params
OpenCV DenseNet PyTorch ResNet
View Project →
Simulation Apr – May 2025

Autonomous Vehicle Visualization Dashboard

12 FPS dashboard overlaying lanes, vehicles, pedestrians, and collision highlights on Tesla Model S streams. 3D classification, pose estimation, and motion segmentation.

12 FPS real-time overlay Full collision prediction
YOLO RAFT OpenPose Blender
View Project →
Simulation Jan – Feb 2025

Physics-Informed Neural Network for Motion

PINN models for predicting object motion under manipulation — incorporating physical priors and engineered features for push/pluck manipulation scenarios.

Improved sim fidelity Physics-constrained
TensorFlow PINN Python
View Project →