Debugging OpenVLA on LIBERO — From 0% to 99% Success
A systematic integration study of OpenVLA and OpenVLA-OFT on the LIBERO manipulation benchmark, isolating eight bugs that silently drove success from 0% to 51.6% (base) and 99.2% (OFT) across 500 evaluation episodes.
Aug 27, 2026
Autonomous Race Driving with Action Mapping RL
Evaluated DDPG and SAC with Action Mapping on a custom autonomous racing simulator using a friction-constrained vehicle model, comparing them to the TD3-AM benchmark from literature.
Apr 10, 2025
Sparse 3D Reconstruction and Bundle Adjustment
Built a custom Structure-from-Motion pipeline using CLAHE, SIFT, triangulation, and GTSAM-based bundle adjustment on a 24-frame dataset of a Buddha sculpture.
Nov 7, 2024
Photo Mosaicking of Low-Contrast Underwater Images
A comprehensive feature-matching and mosaicking pipeline for underwater images using CLAHE, SIFT, RANSAC, affine transformation, and GTSAM-based pose graph optimization.
Oct 5, 2024
Kimera VIO on EuRoC and Custom Datasets
Deployed Kimera-VIO in a reproducible Docker environment, reproduced its EuRoC V1 benchmark, then ran it on a self-collected RealSense D455 indoor sequence — where starting the camera at rest turned out to matter more than any parameter.
Apr 26, 2024
Semantic Geometric SLAM (SG-SLAM) in Dynamic Scenes
Reimplemented SG-SLAM — an RGB-D SLAM system that filters dynamic features by scaling their epipolar residual with a per-class semantic weight — and reproduced its benchmark against ORB-SLAM2 on 14 TUM and Bonn sequences.
Apr 19, 2024
Automated Insect Leg Labeling using DeepLabCut
Built an automated insect leg labeling pipeline using image preprocessing, Shi-Tomasi corner detection, and KMeans clustering to reduce manual labeling effort in DeepLabCut.
Apr 15, 2024
Dead Reckoning with IMU & GPS in Vehicle Navigation
Instrumented a car with a VN-100 and GPS, calibrated the magnetometer for hard and soft iron distortion, fused heading through a complementary filter, and dead-reckoned a 2–3 km Boston drive — holding 2 m for the first 150–200 s before drifting 400 m.
Apr 9, 2024
Digit Recognition & Transfer Learning with CNNs
Built and analyzed a convolutional neural network (CNN) for handwritten digit recognition using MNIST, then applied transfer learning to recognize Greek letters and conducted real-time video-based classification.
Apr 1, 2024
Real-Time 2D Object Recognition with Feature Matching
A real-time C++/OpenCV recognition system built from first principles — 2-means thresholding, hand-written image moments, and shape features that separate 11 object classes without a neural network in sight.
Mar 30, 2024