Projects
Jump to: Motion Planning & Control | Robotics & Perception | Computer Vision & ML | Interdisciplinary
🤖 Motion Planning & Control
Research-level projects in model predictive control, SLAM, and bio-inspired mechanism design.
STORM: Stochastic Model Predictive Control Framework for xArm7 Manipulator
Adapted and deployed the open-source STORM MPC framework on the 7-DOF xArm7 robot to enable real-time joint-space motion planning and control at 100 Hz. Implemented GPU-accelerated trajectory sampling using PyTorch for massively parallel rollout evaluation, and designed feedback-driven optimization for reactive pose tracking and dynamic obstacle avoidance.
Modified robot-specific kinematic models, integrated ROS control interfaces, and implemented Kalman filtering for robust state estimation and smooth motion generation in cluttered environments.
Modified robot-specific kinematic models, integrated ROS control interfaces, and implemented Kalman filtering for robust state estimation and smooth motion generation in cluttered environments.
Impact: Achieved <10ms compute latency for real-time replanning, enabling the manipulator to reactively adapt to moving obstacles while maintaining trajectory optimality.
📂 Code 🎥 Demo (Coming Soon)Pose Graph Optimization for 2D SLAM
Built a pose graph-based backend for 2D SLAM using the g2o optimization framework, integrating odometric constraints and loop closure detections to refine robot trajectory estimates. Implemented robust M-estimators (Huber kernel) to handle outlier measurements and improve convergence stability.
Evaluated performance against ground truth on benchmark datasets, achieving 40% reduction in trajectory error compared to odometry-only baseline through graph-based error minimization.
Evaluated performance against ground truth on benchmark datasets, achieving 40% reduction in trajectory error compared to odometry-only baseline through graph-based error minimization.
Impact: Demonstrated the importance of loop closure integration for long-term localization accuracy in robotics mapping applications.
📂 Code 📊 ResultsTheo Jansen Walking Mechanism
Designed and fabricated a planar walking mechanism based on the Theo Jansen linkage—a biomimetic leg design known for smooth, energy-efficient gait patterns. Developed a forward kinematic model in Python to simulate foot trajectory and optimize gait parameters before physical implementation.
Fabricated the linkage using laser-cut acrylic components and programmed Dynamixel servos for coordinated multi-leg motion control. The project integrated analytical modeling, CAD design (SolidWorks), and embedded systems programming to achieve stable, adaptive locomotion on varied terrain.
Fabricated the linkage using laser-cut acrylic components and programmed Dynamixel servos for coordinated multi-leg motion control. The project integrated analytical modeling, CAD design (SolidWorks), and embedded systems programming to achieve stable, adaptive locomotion on varied terrain.
Impact: Successfully demonstrated bio-inspired robotics design principles with hardware validation, bridging simulation and real-world deployment.
📂 Code 📊 PresentationRobotics Planning & Navigation (RPN)
Robotics Planning & Navigation (RPN)
Implemented advanced planning algorithms for holonomic and non-holonomic robots in 2D PyBullet simulation environments:
📂 Code
- RRT (Rapidly-exploring Random Tree): Developed probabilistic path planning for safe navigation around dynamic and static obstacles with collision checking.
- Model Predictive Control (MPC): Designed MPC using the MPPI (Model Predictive Path Integral) update rule for unicycle kinematics under velocity, acceleration, and lane-keeping constraints.
- Bernstein Polynomial Trajectories: Modeled smooth, differentiable trajectories for non-holonomic robots using Bernstein basis functions, incorporating obstacle avoidance and waypoint optimization via the CEM (Cross-Entropy Method) update rule.
📂 Code
👁️ Robotics Perception & Sensor Fusion
Stereo Dense Reconstruction
Developed a dense 3D reconstruction pipeline using OpenCV's Semi-Global Block Matching (SGBM) algorithm to compute disparity maps from stereo image pairs. Generated detailed colored point clouds with Open3D for robotic scene understanding and navigation in 3D environments.
Applied WLS (Weighted Least Squares) filtering to refine disparity maps and reduce noise, enabling accurate depth perception for autonomous manipulation and obstacle avoidance tasks.
📂 Code
Applied WLS (Weighted Least Squares) filtering to refine disparity maps and reduce noise, enabling accurate depth perception for autonomous manipulation and obstacle avoidance tasks.
📂 Code
Camera–IMU Calibration
Developed a precise camera–IMU extrinsic calibration pipeline using ROS2 and Intel RealSense D435i sensors. Estimated the 6-DOF transformation between camera and IMU frames to enable tightly-coupled sensor fusion for visual-inertial odometry (VIO).
Achieved <2° orientation MSE through batch optimization of synchronized image-IMU measurements, improving motion tracking accuracy for mobile robot localization.
📂 Code
Achieved <2° orientation MSE through batch optimization of synchronized image-IMU measurements, improving motion tracking accuracy for mobile robot localization.
📂 Code
🔍 Computer Vision & Machine Learning
Enhancing Weak Planetary Signals in High-Contrast Exoplanet Imaging using Iterative RPCA
Developed a novel iterative Robust Principal Component Analysis (RPCA) model integrating SVD-based low-rank decomposition and Angular Differential Imaging (ADI) for separating faint planetary signals from overwhelming stellar noise in high-contrast imaging data.
Applied the method to ESO–VLT SPHERE observations of β Pictoris b, achieving uniform SNR enhancement across all angular separations and significantly reducing quasi-static speckle noise that typically dominates direct exoplanet detection.
Applied the method to ESO–VLT SPHERE observations of β Pictoris b, achieving uniform SNR enhancement across all angular separations and significantly reducing quasi-static speckle noise that typically dominates direct exoplanet detection.
Impact: This approach improves the detectability of faint exoplanets in archival data, enabling reanalysis of existing datasets for new discoveries without additional observing time.
📂 CodeComputer Vision Course Projects
Computer Vision Course Projects
Explored classical and modern computer vision approaches across multiple problem domains:
📂 Code
- Image Classification: Implemented handwritten digit recognition on MNIST using classical methods (SIFT + Bag of Visual Words + SVM) and deep learning (CNNs, Vision Transformers). Analyzed hyperparameter effects on accuracy and data efficiency.
- Detection & Tracking: Developed Viola–Jones face detector with IoU-based multi-frame tracking; fine-tuned YOLOv8 on Open Images dataset, comparing pretrained vs. scratch training and varying dataset sizes.
- Semantic Segmentation: Trained U-Net models on Cityscapes dataset, investigating the impact of skip connections on mean IoU performance and convergence speed.
- Multimodal Learning: Compared CLIP (Contrastive Language–Image Pretraining) and ImageNet-pretrained ResNet-50 encoders for zero-shot classification; evaluated FP16 mixed-precision inference for speedup and memory savings.
📂 Code
Digital Image Processing — Classical Algorithms
Digital Image Processing
Implemented classical image processing algorithms from scratch to build foundational understanding of spatial, frequency, and feature-space techniques:
📂 Code
- Core Operations: Pixel-level brightness/contrast adjustments, grayscale ↔ pseudocolor conversion, video frame manipulation.
- Histogram Processing: Custom histogram equalization; bit-depth quantization; LSB steganography for encryption/decryption.
- Filtering & Edge Detection: Manual implementation of mean, median, Gaussian, bilateral, Sobel, Prewitt, Roberts, and Laplacian filters; spatial vs. frequency domain analysis.
- Morphological Processing: Segmented overlapping coins; signature verification via custom erosion, dilation, and skeletonization.
- Feature Detection: Otsu and adaptive thresholding; Hough line/circle detection; Harris corner detection for geometric feature extraction.
📂 Code
SMAI (Statistical Methods in AI)
Statistical Methods in Artificial Intelligence | IIIT Hyderabad
Developed ML and deep learning pipelines from scratch and using libraries, demonstrating strong understanding of statistical learning theory, model evaluation, and neural architectures:
📂 Code
- Implemented Decision Trees, kNN, Naive Bayes, GMMs from scratch with custom clustering metrics.
- Built PCA-based Eigenface Recognition system for face classification.
- Trained CNNs and Autoencoders on CIFAR-10 and MNIST; developed RNN–LSTM networks for IMDB sentiment analysis.
- Kaggle Mini-Project: Age prediction from facial images using VGG16 transfer learning, achieving top leaderboard performance.
📂 Code
Additional AI & ML Projects
Megathon – Social Media Personality Analysis
Developed an Android application that analyzes social media activity and psychometric test results to classify users as better suited for tech or sales roles. The app outputs percentage match scores for each category, aiding recruiters in candidate shortlisting.
📂 Code
📂 Code
Movie Recommender System
Built a movie recommender system using collaborative filtering and matrix factorization (SVD-based) to provide personalized recommendations based on user rating patterns and preferences.
📂 Code
📂 Code
Precog Research Projects — Election Analysis
Contributed to political data analysis projects under the Precog research group at IIIT Hyderabad:
- Election Classification: Implemented classification models to categorize election-related social media posts for large-scale political discourse analysis.
- Election Spends Analysis: Analyzed and classified election advertisement data to estimate campaign expenditures using multimodal (text + image) analysis pipelines.
🧠 Interdisciplinary Research
Exploring cognitive neuroscience and behavioral research—skills transferable to human-robot interaction, AR/VR interfaces, and decision-making models.
VR & Spatial Memory (Cognitive Neuroscience)
Designed a between-subjects controlled experiment investigating how immersive VR environments vs. real-world settings affect spatial memory encoding and retrieval. Participants (ages 18–40) were trained and tested across three complexity levels: object placement, route navigation, and maze navigation.
Measured recall alignment error, response times, and hypothesized brain activation patterns (hippocampus, medial entorhinal cortex). Results demonstrated superior spatial recall performance in real-world encoding, highlighting sensorimotor engagement's critical role in memory consolidation and revealing VR's limitations in complex spatial tasks.
Measured recall alignment error, response times, and hypothesized brain activation patterns (hippocampus, medial entorhinal cortex). Results demonstrated superior spatial recall performance in real-world encoding, highlighting sensorimotor engagement's critical role in memory consolidation and revealing VR's limitations in complex spatial tasks.
Relevance to Robotics: Understanding human spatial cognition informs the design of AR/VR-based robot teleoperation interfaces and human-robot collaboration in shared physical-virtual environments.
📄 ReportGender Differences in Reward-Based Effort
Replicated and extended a Nature study on gender differences in cost-benefit decision-making by analyzing effort allocation across 81 participants. Applied advanced statistical methods including mixed-effects models, Kruskal-Wallis tests, and post-hoc pairwise comparisons.
Confirmed that vagus nerve stimulation (VNS) significantly enhances reward-driven motivation (p < 0.001) and demonstrated that gender independently modulates effort maintenance (Cohen's d = 0.338). Provided new insights into the interplay between physiological interventions and behavioral decision-making processes.
Confirmed that vagus nerve stimulation (VNS) significantly enhances reward-driven motivation (p < 0.001) and demonstrated that gender independently modulates effort maintenance (Cohen's d = 0.338). Provided new insights into the interplay between physiological interventions and behavioral decision-making processes.
Relevance to Robotics: Understanding effort allocation and motivation models can inform reward function design in reinforcement learning for autonomous systems and human-robot collaboration frameworks.
📂 Code 📄 Report