Andres Robles Gil

Water Pollution Analysis with Machine Learning

This project is a comprehensive analysis of water pollution trends using advanced data science techniques and machine learning algorithms. By examining real-world datasets, we sought to understand the key factors that drive pollution patterns across different regions.

Data Engineering and Processing:

• Data Acquisition: Sourced large water pollution datasets from public repositories, including historical water quality records.
• Data Cleaning: Removed inaccuracies, handled missing values, and eliminated outliers.
• Feature Engineering: Generated new features to enhance the dataset’s predictive capabilities, such as pollution growth rates and seasonal fluctuations.
• Data Pipelines: Built automated data pipelines using Python libraries such as pandas and scikit-learn.
• Data Visualization: Visualized water quality metrics and pollution levels using matplotlib and seaborn.

Modeling Techniques:

• Linear Regression: Predicted future pollution levels based on historical trends and regional data.
• K-Means Clustering: Identified distinct pollution regions and high pollution zones.

Key AI/ML Contributions:

• Developed a predictive model using regression techniques with Python and scikit-learn.
• Applied unsupervised clustering to identify pollution trends across different regions.
• Deployed the model using Jupyter notebooks for replication of results.

View Water Pollution Project on GitHub

Thesis: Synthetic Data Generation for Grasp Quality Evaluation

This thesis focuses on generating synthetic data for evaluating grasp quality in robotic bin-picking tasks. We developed a data generation pipeline that utilizes Blender through its Python API for creating synthetic 3D environments, simulating various grasp conditions. The synthetic data was used to train a neural network model for predicting grasp quality.

Data Engineering and Processing:

• 3D Scene Creation: Built synthetic 3D environments using Blender to simulate bin-picking scenarios for robotic grasping. The scenes included random object placements, lighting conditions, and background variations to ensure diverse data generation.
• Grasp Simulation: Used Isaac Sim via its Python library to simulate grasping tasks and evaluate robotic grasp performance. The simulator provided detailed physics-based grasp trials that were further analyzed for training data generation.
• Grasp Quality Evaluation: Generated grasp quality maps based on object geometry, surface smoothness, and other physical factors. These maps were used to label the data and assess grasp success or failure under different conditions.
• Data Labeling: Labeled the generated data for training machine learning models to predict optimal grasp positions, accounting for variables such as object orientation, contact points, and surface friction.

Key AI/ML Contributions:

• Developed a custom Blender add-on through its Python API for generating large-scale synthetic data specific to robotic bin-picking scenarios.
• Implemented Isaac Sim simulations to evaluate grasp efficiency under varying conditions, enabling a more realistic dataset for grasp quality prediction.
• Deployed machine learning models using TensorFlow and a Hugging Face UNet model to predict optimal grasp points based on synthetic data labeled from simulations.
• Utilized Google Cloud for large-scale data generation and model training, leveraging its compute resources to scale simulations and accelerate training times.
• Achieved over 95% grasp success rate across diverse and challenging environments, proving the effectiveness of synthetic data in robotic grasp applications.

View Thesis Project on GitHub