AI

Ensemble Convolutional Neural Networks for Robotic Planar Grasping

Overview

This project addressed the challenge of generalized robotic grasping, where robots must reliably grasp unknown objects in uncontrolled environments. Modern approaches like Best Candidate selection, Grasp Quality Estimation (e.g., DexNet-4.0), and Generative methods (e.g., GGCNN) struggle with generalization due to sensitivity to object shape, lighting, and environment.

Testing and Results

We proposed an Ensemble Convolutional Neural Network (ECNN) using a Mixture of Experts model. Multiple grasping networks (experts) were combined with a gating network that assigns input-dependent weights, allowing the ensemble to exploit each expert’s strengths and reduce generalization errors. Three ensemble variants were developed: Constant Weights, Image-based (ImECNN), and Grasp+Image-based (GrImECNN). The networks were trained using the Cornell Grasping Dataset and tested on real-world YCB objects with a Franka Emika robot. Results showed up to a 6% increase in classification accuracy compared to the strongest individual expert, demonstrating improved grasp quality estimation with minimal computational overhead.

Report

Download Report

Machine Learning

Just NERF It!

I collaborated with a team to develop a system that converts a small set of photos into immersive 3D scenes using Neural Radiance Fields (NeRF). By capturing multi-angle images and estimating camera poses with COLMAP, we reconstructed scenes that could be viewed from new perspectives. The system performed well in static environments, even capturing reflections, though real-world scenes required more coverage than synthetic ones. To expand functionality, we explored integrating object labels, enabling users to directly edit scenes— inserting, removing, or swapping 3D objects for interactive use cases.

Report