Enhancing Diffusion Policy Generalization through Structured Visual Representations

An attention-based masking encoder (Attn-DP) trained end-to-end under BC loss, improving OOD generalization by 12.4% across 12 unseen YCB objects. Integrates frozen SAM3 at inference for zero-shot segmentation without fine-tuning. The Attention U-Net learns to suppress task-irrelevant background from imitation losses alone.

RoboMimic2D: Lightweight 2D Manipulation Benchmarks for Policy Learning

A Gymnasium-compatible 2D simulation benchmark mirroring RoboMimic's task structure. Physics-based grasping via Pymunk with a gripper state machine — grip force computed from friction coefficients, not hardcoded. Zarr-backed episodic replay buffer for imitation learning data collection.

Fuzzy-based Techniques for Clustering in Wireless Sensor Networks

Artificial Electric Field Algorithm: A Python Package for Global Optimization

A pip-installable Python package implementing the Artificial Electric Field optimization algorithm. Converges to global minima/maxima for multi-dimensional, non-linear problems with 2D convergence visualization.

Augmented Reality C++ Pipeline

Real-time AR in C++ with dynamic camera calibration, markerless target tracking via ORB feature matching and RANSAC homography, and an OpenGL/GLFW rendering pipeline for shaded 3D geometry overlays on live video.

GELLO Teleoperation & Real-Robot Data Collection

Built a GELLO teleoperation device from scratch for UR5 arm demonstration collection. Reproduced Diffusion Policy baselines across RoboMimic and RoboCasa benchmarks. Benchmarked UNet, ViT, and foundation-model visual encoders (DINO, SAM, VLMs) for manipulation policy learning.

No-Reward Long-Horizon Planning with Latent World Models

Intrinsic Curiosity Module (ICM) with Deep Q-Network for sparse-reward environments. 75% success rate increase over vanilla DQN on long-horizon FrozenLake env.

Brain-Controlled Robot Arm via EEG Calibration

Controlling a Hugging Face Lerobot arm using a Muse 2 EEG headset by Interaxon. A simple blink translates into physical movement — with zero data collection, zero fine-tuning, and zero pretrained models. Specific facial movements are tied to specific robot joint movements in this project. The setup relies entirely on pure device calibration mapped directly to brain signals, taking only seconds to set up.

Physics Simulations with Pymunk & Pygame

A collection of 2D physics simulations with dual implementations: raw numpy and Pymunk-backed. Documented in a three-part technical blog series.