Jorge Ramirez

• State estimation & modeling — modeled vehicle kinematics with a bicycle model and fused GPS/IMU data through a dead-reckoning and heading-corrected EKF for robust pose estimates.
• Control — closed the loop with longitudinal speed control and Stanley-based geometric steering for trajectory tracking.
• Perception (vision) — calibrated cameras for undistorted imagery, then layered lane detection, object detection (YOLOv8), and bird's-eye-view lane-keeping with pure-pursuit control.
• Perception (LiDAR) & planning — built occupancy grid maps from LiDAR for environment interpretation, and planned global paths with A*.

[Add results summary here — e.g. tracking accuracy, detection performance, or a link to the demo run.]

Compared A* and Hybrid A* path planning for autonomous parking in CARLA (Town05). Both planners were sourced from an open-source motion planning library (zhm‑real) — my contribution was the system integration: building the CARLA obstacle pipeline from vehicle bounding boxes, a custom path follower with crash detection (CARLA's built-in manager doesn't support custom path tracking), Stanley controller tuning to match the Audi e-tron model, and the trial evaluation framework. This closed-loop setup — planning → tracking → crash detection — is what caught A*'s Trial 1 collision, directly demonstrating the kinematic-feasibility gap the project set out to test.

Hybrid A* discretizes heading into 15° bins and samples 39 steering angles per expansion, propagating a bicycle kinematic model and using Reeds‑Shepp shortcuts to terminate search early once a direct connection to the goal is feasible.

Five trials with varying obstacle layouts in a 57×34m parking lot, isolating the planner as the only variable between runs to attribute outcome differences directly to the algorithm.

Co-developed with Vanessa Guzman — my contributions: simulation setup, obstacle pipeline, CARLA integration, path follower, controller tuning.

Built a vision-guided pick-and-place system on an Open Manipulator‑X arm, using an Intel RealSense camera to detect colored spheres via Hough Transform circle detection and HSV-based color classification. Ball positions were localized by repurposing an AprilTag PnP pipeline to get camera-frame coordinates, then transformed to robot frame using a calibration matrix from an earlier lab. Each pick used four chained cubic trajectories (approach → descend → retract → deliver to bin), with the gripper timed to close and open at the grasp and drop points.

• Hough Transform over contour analysis — justified by known, well-defined circle geometry, despite the higher memory cost.
• Cubic trajectories over PID control — ball positions were static, so smooth open-loop trajectories were sufficient; PID would only have been necessary for a dynamic environment.

Image pipeline: HSV conversion → grayscale + CLAHE contrast enhancement → Gaussian blur → Hough circle detection → HSV-hue color classification, with workspace-bounds filtering to reject false-positive detections outside the robot's reachable area.

Collaborative lab with Anthony Man — shared design and implementation across CV, kinematics, and control.