I am an MS Mechatronic Engineering student working on:

My research focuses on developing, training, testing and eventually deploying DRL-based quadrotor navigation policies using a simulation-first robotics pipeline.

• Deep Reinforcement Learning for autonomous drone navigation
• Flightmare-based quadrotor simulation
• PyBullet drone RL baselines
• ROS 2-based robotics integration
• RealSense-based perception
• Jetson Orin Nano edge deployment
• Sim-to-real transfer for aerial robots

• Ubuntu 22.04 native Linux
• Python virtual environments
• PyTorch
• Gymnasium
• Stable-Baselines3
• RLlib / Ray
• PyBullet drone simulation
• Flightmare
• ROS 2 Humble
• Gazebo Fortress
• RealSense RGB-D / VIO sensors
• Jetson Orin Nano deployment target

1. Validate DRL algorithms on simple Gymnasium environments
2. Test drone control baselines in PyBullet
3. Use Flightmare as the main thesis simulator
4. Build a modern Gymnasium-compatible Flightmare wrapper
5. Train policies using SB3, RLlib and PyTorch
6. Integrate policy inference with ROS 2
7. Test RealSense-based perception
8. Move toward Jetson-based deployment

My long-term goal is to develop reliable DRL-based drone navigation policies that can be trained in simulation and later adapted for real-world autonomous flight.

• drl-drone-navigation-thesis
• flightmare-gymnasium-wrapper
• pybullet-drone-rl-baselines
• ros2-drone-policy-inference
• realsense-ros2-drone-perception
• autonomous-drone-racing-literature

My MS thesis/research direction is:

Deep Reinforcement Learning-Based Autonomous Drone Navigation

Current focused technical direction:

From Privileged State to Vision Features: A Practical Flightmare Pipeline for DRL-Based Drone Gate Navigation

The portfolio is organized as a staged research pipeline.

Main public research hub for my MS thesis project.

This repository tracks the research pipeline, simulator decisions, setup notes, progress logs, and future experimental roadmap.

Foundational RL and PyBullet drone baseline repository.

This repository documents early reinforcement learning algorithm exploration using Gymnasium, Stable-Baselines3, RLlib/Ray, PyTorch, and PyBullet.

Current highlights:

• SB3 CartPole-v1 experiments with PPO, A2C, and DQN
• RLlib CartPole-v1 experiments with PPO and DQN
• SB3 Pendulum-v1 continuous-control experiments with PPO and SAC
• SB3 MountainCarContinuous-v0 experiments with SAC, TD3, and PPO
• Initial external gym-pybullet-drones visual simulation testing
• Continuous-control learning path toward drone RL

Reusable headless RL wrapper around Flightmare.

This repository modernizes Flightmare headless RL training using Python 3.10, Gymnasium-style wrappers, Stable-Baselines3, and PyTorch.

Current highlights:

• Raw Flightmare vector wrapper
• Gymnasium-compatible single-env wrapper
• SB3-compatible VecEnv wrapper
• Scaled-action VecEnv wrapper
• PPO headless training and evaluation
• Action-scale ablation

Flightmare UnityBridge visualization and racing-track tooling.

Current highlights:

• YAML track loader
• Scene selection for Warehouse, Industrial, Garage, and NatureForest
• Local-origin coordinate calibration
• RGB/depth/segmentation camera capture testing
• Gate-passing CSV logger
• UZH/RPG-inspired SplitS, Figure 8, and Kidney track templates
• GIF previews for visual demonstration

Reusable Flightmare vision-aware DRL gate-navigation pipeline.

This repository contains implementation-oriented tools for:

• Privileged-state teacher PPO training
• Teacher rollout dataset generation
• Compact gate/vision-feature observations
• 25D vision-proprioceptive student observations
• Imitation-learning student policy training
• PPO-from-scratch baselines
• IL-initialized PPO fine-tuning
• Robustness evaluation
• UnityBridge replay export

This repository shows the engineering implementation and reusable research pipeline.

Curated Flightmare visual evidence and result archive for DRL-based autonomous drone gate navigation.

This repository documents the progression from privileged PPO teacher training to teacher rollout generation, imitation learning, PPO-from-scratch baselines, robustness evaluation, and UnityBridge replay validation.

Highlights:

• Phase-wise thesis reports
• Policy comparison tables
• Failure-mode analysis
• Observation-space comparison
• Phase 5 imitation-learning diagnostics
• Phase 8 robustness plots
• Phase 9 UnityBridge replay GIFs and MP4s
• Onboard RGB and feature-camera consistency evidence

Primary result:

A 25D vision-proprioceptive imitation-learning student completed the T02 three-gate Flightmare replay in UnityBridge visual validation.

Boundary:

This is replay-based visual validation of a compact vision-proprioceptive policy, not raw RGB end-to-end real-drone deployment.