Diffusion-transformer policy for 26-DOF dexterous pick-and-place. Maps dual-view RGBD and proprioception to action chunks with SE(3) geometric regularization. Includes Vision Pro-based human retargeting in MuJoCo.

6-DoF variable-mass lander in MuJoCo with 2-DoF gimbaled thrust; PID baseline vs PPO policy; pose-targeted touchdown and fuel-efficiency metrics.

Monocular depth pipeline for human-robot walking environments: ZoeDepth baseline, metric reprojection to 3D point clouds, and iterative RANSAC to extract stair planes and dimensions.

Compared tabular Q-learning, DQN, PPO, and SAC on noisy Cart-Pole; designed noise-robust wrappers, tuned hyperparameters, and analyzed robustness under disturbances and varying noise levels.

Virtual Holonomic Constraints with polynomial gait design, feedback linearization, and terrain generalization to slopes and stairs.

Batch Gauss-Newton fusion with EKF; compared batch vs sliding-window estimators for trajectory accuracy.

MATLAB, SolidWorks — Designed a 3D-printed Arduino-based mobile robot in Fusion 360; implemented localization and obstacle-avoidance for maze navigation in MATLAB; commanded the robot via Bluetooth.

ESP32-based driver with SPI multiplexing; API control; interactive 3D Snake and Pong demos.

Gmapping, frontier exploration, PID control; OpenCV perception and a CNN for 7-class emotion recognition.

Planetary gearbox quadrupling torque at >90% efficiency; kinematics, stress, thermal, and vibration analyses.