Luca's Engineering Portfolio

Welcome to my GitHub portfolio!
Here, I showcase my work completed during university and personal projects.

PI-Bot: Distributed 3-DOF Robotic Arm (Personal Project)

This repository contains the full source code and hardware configuration for the PI-Bot, a custom-built 3-DOF robotic arm. Developed as an end-to-end solo project, it demonstrates the integration of distributed software systems, hardware control, and a hybrid robot control system. The system is built on a distributed ROS2 architecture, where a workstation handles high-level motion planning and a Raspberry Pi 4 manages the low-level hardware abstraction layer.

System Demo

Technical Highlights & Implementation

This project covers the entire robotics stack, from mechanical coordination to high-level software orchestration:

• Distributed ROS2 Architecture: Engineered a multi-node system where computation is split between a PC (running MoveIt2 & GUI) and a Raspberry Pi 4 (Servo/ADC nodes), communicating seamlessly via ROS2 Topics and SSH.
• Hybrid Control System: Developed two distinct operational modes:
  • GUI-Control: Asynchronous trajectory execution using MoveIt2 and PySide6.
  • Hardware-in-the-Loop (HIL): Real-time manipulation via analog potentiometers, allowing the user to "teach" the robot movements.
• Custom GUI (PI-Bot Control Center): Designed and implemented a dashboard using PySide6 (Qt). Key features include a Sequence Recorder to store/execute movement patterns and real-time state synchronization to prevent hardware jumps.
• Low-Level Hardware Interfacing: Developed custom nodes for the ADS7830 ADC (I2C) and PCA9685 PWM (I2C) to bridge the gap between analog sensors and the digital ROS2 environment.
• Kinematics & Digital Twin: Created a precise URDF model and configured the TF-tree for real-time visualization in RViz, ensuring the digital twin perfectly reflects the physical state of the 3-DOF arm.
• Motion Planning: Integrated the MoveIt2 framework for inverse kinematics (IK) and collision-free path planning, utilizing custom service interfaces for cartesian positioning.

---

Autonomous Driving Project (University Project)

This repository contains the documentation and source code for our autonomous driving system, developed as part of a university course on Introduction to ROS during the summer term of 2025.

Our team engineered a comprehensive solution for self-driving within a simulated urban environment. The primary objective was to build a robust system capable of real-time navigation, traffic light compliance, and collision avoidance using ROS (Robot Operating System) and a Unity-based simulator.

System Demo

Key Learnings & Project Highlights

This project provided invaluable hands-on experience and a deep dive into advanced robotics and AI concepts:

• ROS (Robot Operating System): Gained a strong understanding of ROS architecture, including node communication, topics, services, and the orchestration of complex robotic systems via launch files.
• Autonomous Driving Concepts: Practical application of fundamental algorithms for perception in a simulated environment.
• Sensor Data Fusion: Developed skills in processing, aligning, and fusing data from depth images and cameras to construct a comprehensive environmental model.
• Machine Learning for Perception: Acquired practical experience in fine-tuning and deploying YOLOv8 networks for real-time object detection (specifically cars and traffic lights) in simulated scenarios.
• Git & Version Control: Mastered collaborative codebase management using Git, including efficient handling of large binary files with Git Large File Storage (Git LFS) and maintaining a clean, effective project history.
• Team Collaboration: Successfully navigated the complexities of agile team development, clearly defining responsibilities, and seamlessly integrating individual software modules into a cohesive system.

My Contribution (Luca Brenzinger)

• Perception Module (Lead): I led the design and implementation of key components within the perception pipeline. This included developing the static_tf_ins_to_cameras node for precise sensor frame transformations and configuring the pointcloud.launch file for robust 3D point cloud generation.
• YOLOv8 Network Fine-Tuning & Data Augmentation: I made significant contributions to the fine-tuning of the YOLOv8 object detection model. My work involved setting up and executing data augmentation strategies to enhance model generalization and training the network on custom simulation data for accurate car and traffic light detection.
• Object Detector Node Development: I actively developed and refined the object_detector node, which continuously processes incoming images to identify objects of interest, publishing their bounding box coordinates and class labels for downstream modules.
• 3D Car Position Estimator Node: My involvement extended to the car_position_estimator node, where I contributed to estimating 3D world positions of detected cars through depth-based triangulation and visualizing these results in RViz.
• Documentation: I contributed to the comprehensive project documentation, providing detailed explanations of key modules and their functionalities to ensure clarity and maintainability.

---

AI for Industry Challenge (Intrinsic Team Competition)

This repository contains our solution to Intrinsic's AI for Industry Challenge: a robot arm autonomously plugging SFP and SC network connectors into their ports. Competing as a two-person team against groups of up to ten, we built two complete, very different insertion pipelines over two rounds — a fully self-trained vision pipeline for Qualification, then a force-controlled policy for Phase 1 once port detection was handled by Intrinsic's FlowState.

System Demo

phase1-sc-insert.mp4 SC insertion

phase1-sfp-insert.mp4 SFP insertion

Results

Round	Approach	Result
Qualification	Own vision pipeline (YOLO port detection + triangulation)	27 / 160 teams advanced
Phase 1	Intrinsic FlowState for perception, force-controlled insertion	14 / 160 teams

Key Highlights

• Custom Vision Pipeline (Qualification): With no port detection provided, we labeled our own dataset and trained a YOLOv8 keypoint model to find each port's corners, then triangulated the multi-camera detections into a 3D port pose.
• Residual Offset-Correction Model: Trained a vision-based regressor on data from our own capture pipeline to correct residual pose error right before insertion — reused unchanged across both rounds.
• Force-Controlled Insertion Policy (Phase 1): Reworked the policy around force control after perception moved to FlowState — detecting contact, distinguishing a clean insertion from an edge catch, and recovering from snags without any camera input.
• Spiral Search Strategy: A two-stage spiral search under soft stiffness for final alignment once the plug is near the port opening, used as a fallback in both pipelines.

---

C++ Autonomous Cruise Control Simulator (Personal Engineering Project)

This repository features a C++ based simulation of a vehicle's cruise control system, leveraging an object-oriented architecture to implement a discrete PID controller.

Figure 1: Shows the simulated step response (0 to 20 m/s) with tuned parameters ($K_p = 5.0, K_i = 0.1, K_d = 0.5$).

It originated as a group project at the Technical University of Munich (TUM). I am currently refactoring and extending the codebase to deepen my C++ knowledge and improve the software architecture beyond the original academic submission. s

Key Highlights

• Object-Oriented Implementation: Clean separation of vehicle dynamics and PID control logic using an OOP approach.
• CMake Build System: Professional project structure ensuring easy compilation and cross-platform compatibility.
• Simulation & Visualization: Full simulation pipeline with CSV data export and Python-based plotting for performance analysis.