Projects
Robotics & Reinforcement Learning
Hypernetwork-based Optimizer
Tech Stack: Python, PyTorch, Gemma LLM, LoRA, Flash Attention 2
- Developed semantically enhanced Hypernetwork-based Optimizers by fine-tuning the Gemma LLM with Distributed Data Parallelism on 4xA100 GPUs.
- Achieved convergence on 10+ numerical and Reinforcement Learning domains using Supervised Fine-Tuning (SFT) on a 1.6B token dataset.
- Accelerated LLM fine-tuning by 50% through Quantization, LoRA, and reasoning distillation, integrating Flash Attention 2 for a 2x faster evaluation pipeline.
FORGE: Program Synthesis for Robotic Manipulation
Tech Stack: Python, LLMs, MuJoCo, UR5 Manipulator
- Co-developed an LLM-based program synthesis framework for compliance-aware robotic manipulation without requiring gradient updates.
- Achieved 1.6 mm peg-in-hole accuracy with a UR5 robot arm by generating code-as-policy.
- Presented at the Southwest Robotics Symposium (SWRS) 2025.
ProPS: Prompted Policy Search
Tech Stack: Python, PyTorch, LLMs, Reinforcement Learning
- Co-developed an LLM-based optimizer that incorporates linguistic reasoning for numerical and RL policy optimization.
- Achieved State-of-the-Art (SOTA) results by outperforming PPO and TRPO on 8 of 15 benchmarks.
- Implemented random projection techniques to reduce neural network dimensionality by 70%, facilitating faster convergence for high-dimensional optimization tasks.
Shared Control UR5 for In-Space Welding
Tech Stack: Python, MuJoCo, OpenCV, ArUco Markers, Inverse Dynamics
- Designed a human-in-the-loop simulation for orbital construction, featuring a 6-DOF SpaceMouse controller and force-based inverse dynamics for a UR5 manipulator.
- Implemented a computer vision pipeline using ArUco markers for real-time localization and pose estimation of metallic workpieces.
- Developed semi-autonomous “Grab” and “Weld” modes with power-grip stabilization, enabling precise end-effector alignment in a zero-gravity environment.
Humanoid Quadruped Motion Planning
Tech Stack: Python, MuJoCo, PCA, CMA-ES, Dynamic Movement Primitives (DMP)
- Implemented a latent space planning approach for Unitree G1 robot locomotion, reducing 29 joint degrees-of-freedom to a rank-2 latent space using PCA.
- Generated stable crawl trajectories using Dynamic Movement Primitives (DMPs) optimized via Covariance Matrix Adaptation Evolution Strategy (CMA-ES).
- Demonstrated that low-dimensional latent optimization could effectively synthesize complex, periodic crawling behaviors without full-body policy training.
Autonomous Indoor Navigation
Tech Stack: Python, ROS2, Gazebo, Turtlebot3, Q-Learning
- Designed a discrete state-action space navigation system by segmenting LiDAR data into spatial sectors and zones.
- Implemented a Q-learning algorithm with a custom reward structure (optimizing for distance, direction, and collision avoidance), training the agent over 400 episodes.
- Achieved successful target navigation with reduced collision rates by iteratively refining the policy based on 144 unique state representations.
Robotic Chess Player using Foundation Models
Tech Stack: Python, UR5, Pi0 VLA, Inverse Kinematics
- Prototyped a robotic chess player using a UR5 manipulator and Vision-Language-Action (VLA) models, integrating perception with contact-rich actuation.
- Implemented language-conditioned control by parsing PGN (Portable Game Notation) strings to align high-level chess moves with low-level robotic motion planning.
- Fine-tuned foundation models with inverse kinematics-generated demonstrations to handle precise piece manipulation.
Computer Vision & Multimedia
Self-Supervised Point Tracking in Turbulent Videos
Tech Stack: Python, PyTorch, DINOv2, RAFT, Restormer, QuickTurbSim
- Engineered a dense point tracking system robust to atmospheric distortion by enhancing the DINO-Tracker architecture with RAFT-based optical flow refinement.
- Created a novel benchmark for turbulent video tracking by augmenting the TAP-Vid dataset using QuickTurbSim.
- Developed a pipeline to estimate turbulence strength () based on the temporal displacement of tracked points and integrated Restormer for video stabilization.
Neural Feature Extraction and Semantic Video Retrieval
Tech Stack: Python, ResNet, OpenCv, Scikit-Learn, Streamlit
- Designed a video retrieval engine using ResNet, HOG, and Color Histograms, reducing high-dimensional features into latent semantics via SVD, PCA, and KMeans.
- Implemented relevance feedback mechanisms (Decision Trees and KNN) to iteratively refine search rankings based on user input.
- Performed spectral clustering to visualize semantic groupings of videos, interpreting inter-label similarities using MDS embeddings.
Cross-Geography Generalization for Flood Segmentation
Tech Stack: Python, Deep Learning, Computer Vision, UAV Imagery
- Developed neural network architectures for the segmentation of flooded regions in UAV-captured aerial images.
- Focused on domain generalization to ensure the model performed accurately across different geographical terrains and conditions.
Multiclass Classification and Verification of Online Signatures
Tech Stack: Python, Scikit-Learn, SVM, Signal Processing
- Developed a signature verification tool using Support Vector Machines (SVM) on time-series data to detect forgeries and classify owners.
- Optimized model efficiency by implementing the Ramer-Douglas-Peucker algorithm for feature dimension reduction without information loss.