Sim-to-Real Reinforcement Learning Walking Including Stairs
Closing the sim-to-real gap for quadruped locomotion. PPO policies trained in Genesis with domain randomization, sensor noise injection, and latency modeling deploy directly on a real Unitree Go2. The robot achieves omnidirectional walking and blind stair climbing, using only proprioceptive sensing and a learned per-leg stiffness that adapts compliance in real time.
Extended Kalman Filter SLAM implemented entirely from scratch in C++ and deployed on a real TurtleBot3 Burger. Landmark detection via LiDAR clustering and circle fitting, Mahalanobis-distance data association, and simultaneous pose and map correction reduce localization error by up to 99.7% compared to pure odometry.
Franka Vision-Guided Pick & Place
A complete vision-guided pick-and-place system integrating YOLOv8 object detection, 3D pose estimation, and MoveIt 2 trajectory planning to autonomously grasp and place objects on a Franka Panda robot.
Group project with Robert, Florian, and Aravind
Prompt-to-Pose Grasp Estimation
A natural language-guided grasp planning pipeline using Grounding DINO for object localization, SAM 2 for segmentation, and Contact-GraspNet to generate 6-DoF grasp poses from text descriptions.
Physics-based simulation and analysis of a Jack-in-the-Box mechanism, modeling the hybrid dynamics of spring oscillation combined with discrete impact events.
A real-time vision system that uses OpenCV tracking and Kalman filter prediction to intercept a falling pen with a robotic arm under tight latency constraints.