Projects

Semantic Segmentation for Crack Detection | UG Thesis II, IIT Bombay

Created a purely synthetic dataset using the data generation pipeline while introducing position randomization in the cracks and then augmented the dataset by playing around with the image characteristics. Performed object detection and semantic segmentation on this synthetic dataset by fine-tuning the Detectron-2 model and achieved promising results for both the tasks. 

Synthetic Data Generation Pipeline via Finite Element Simulations | UG Thesis I, IIT Bombay

Wrote a MATLAB-based Finite Element Simulation pipeline using the pdetoolkit for generating temperature plots to mimic heat transfer experiments for detecting cracks in steel plates. Tested multiple vanilla image processing methods such as Hough Transform, Otsu Thresholding and RANSAC as a proof-of-concept. 

EXOTIC: An Exact, Optimistic, Tree-Based Algorithm for Min-Max Optimization 

Under review at SIAM Journal on Optimization (SIOPT) 

Arxiv Link: https://arxiv.org/abs/2508.12479 | GitHub: https://github.com/ChinmayPimpalkhare/EXOTIC

Implemented a new algorithmic solver (EXOTIC) in Julia for solving convex-non-concave optimization problems. Used hierarchical tree-based partitioning and converted the problem into a convex semi-infinite problem.  Benchmarked the method on test examples, outperformed SOTA gradient-based methods like AGP and GDA.

Algorithmic Feedback Synthesis for Ensuring Robust Strong Invariance using Convex Semi-Infinite Programming | Research Project, IIT Bombay

Devised a numerically tractable technique for synthesizing feedbacks to ensure robust positive invariance of given compact sets for nonlinear control systems. Converted the feedback search problem into a linear approximation problem in a functional analytic sense, and verified the 'tangent conditions' for robust positive invariance. Deployed the algorithmic procedure based on the work DACC22. Tested the method on multiple dynamical systems. 

Coverage Maximization for UAV Surveillance on Non-convex Domains | Research Project

Coded the genetic algorithm from scratch in Python and implemented it for maximizing the coverage of a UAV sensor network on domains that are either (a) non-convex (b) disconnected or (c) both. Used Monte Carlo simulations for estimating the area coverage. Devised a proof of concept to simplify complicated domains into polygonal boundaries using the RDP algorithm. 

Work was presented virtually at the 6th National Conference on Multi-Discplinary Optimization and Analysis (NCMDAO 2023) at IIT Guwahati. Paper to be published in Springer Notes in Mechanical Engineering

Rescue Robot Optimization | ME 308, IIT Bombay

Optimized various parts of a rescue mission pipeline. Generated an environment mimicking disasters using Perlin noise for ensuring gradient properties. Implemented a genetic algorithm for optimal sensor mapping in a given terrain. Abstracted the task of human search using drones as a global maximization problem and devised a novel swarm algorithm for tackling multimodal reward functions expressed as a mixture of Gaussians. Implemented linear programming method for allocation and distribution of tasks to robots based on their individual capabilities. 

Monte Carlo Tree Search for playing Billiards efficiently | CS 747, IIT Bombay

Implemented an agent employing Monte Carlo Tree Search for playing Billiards efficiently in scenarios where the cue angle and force inputs have external disturbances and tested it successfully on a pygame simulator. 

Studying Markov Decision Processes through Half-field Football | CS 747, IIT Bombay

Modelled a game of football in a realistic way and abstracted it as a Markov Decision Process. Implemented multiple MDP planning algorithms such as (a) Value Iteration (b) Howard's Policy Iteration (c) Linear Programming and compared their performance against various stationary opponent policies. 

Analysis of Online Algorithms for Multi-Armed Bandits | CS 747, IIT Bombay

Implemented the UCB, KL-UCB and Thompson Sampling Algorithms for cumulative regret minimization in bandit settings. Studied the performance as function of reward means for Bernoulli bandits in (a) a partially observable setting (b) when one of the arms produces noisy outputs that are faulty. 

Code-Switching in Automatic Speech Recognition | CS 753, IIT Bombay 

Used the OpenAI Whisper-small model to recognize code-switched Hindi-English speech. Implemented four different decoding strategies (a) Zero-shot Greedy decoding (b) Constrained Filtering-based Greedy decoding (c) Beam Search decoding (d) Constrained Beam Search decoding and compared their performance. 

Automatic Speech Recognition using the Conformer architecture | CS 753, IIT Bombay 

Learnt how to run a SpeechBrain ASR recipe. Studied the Connectionist Temporal Classification (CTC) loss and implemented a variation of it using hooks in Pytorch to calculate the loss function at intermediate layers instead of just at the last layer. Implemented the Conformer architecture with this modified loss function on an audio dataset.

Facial Emotion Generation using VAE and DDPM | EE782, IIT Bombay

Used the Cleaned Expressions in the Wild (ExpW) dataset to train a Variational Autoencoder to semantically encode the information in the image. We then concatenated the image embedding with the emotion label to produce a conditional label, which was then passed as input to a Denoising Diffusion Probabilistic Model to produce the output image. 

Metric Learning and Generative AI for Facial Imagery | EE782, IIT Bombay

Trained a Siamese Network to predict whether a pair of images belonged to the same person or not using a metric learning scheme. Trained a GAN to synthesize facial images taking a Gaussian noise vector as an input and a facial image as an output.  

Technical Summer School on Natural Language Processing | IIT Bombay 

Performed sentiment analysis using Natural Language Processing (NLP) methods such as Skip-gram and GLoVE. Explored various RNN, LSTM and transformer architectures for language modelling and generation tasks. 

Algorithms for Weakly Supervised Learning | Summer Research Project, IIT Bombay 

Studied and surveyed various algorithms and techniques for weakly supervised learning including (a) Weakly Supervised Deep Detection Networks (b) Multiple Instance Learning (c) Multiple Instance Self-Training (d) Adversarial Distributional Training and (e) Black Box Adversarial Attacks using Evolutionary Strategies

Out Of Distribution Entity Detection using Confidence Thresholds | CS 419, IIT Bombay

Implemented a pre-trained ResNet50 model along with the cleanlab library in Python to generate neural embeddings from images from the CIFAR-10, MNIST and Fashion MNIST datasets. Classified them into in-distribution and Out-of-distribution based on confidence-based thresholding.  

Data Analysis of a Chemical Processing Plant | DS203, IIT Bombay

Performed denoising on the data using moving average smoothing & missing value imputation. Performed exploratory data analysis using Spearman correlation method to handle multicollinearity. We implemented linear regression, random forest and XGBoost regressor models to classify equipment vibration levels based on control parameters. Used Recursive Feature Elimination with Cross-Validation (RFECV) to obtain the optimal number of features for the models. 

Awarded highest possible AP (Advanced Performer) grade, only given to 3 out of 200+ enrolled students 

Time-series Analysis of a Solar Power Generation Site | DS203, IIT Bombay

Performed denoising of the time-series data using moving average smoothing and convolution-based approaches. Extracted time domain features such as peak-to-peak amplitude and trained multiple machine learning models (a) Linear Regression (b) SVM (c) Random Forest to perform data imputation on bad sensor inputs using good sensor inputs.  

Motion Planning and Coordination of Autonomous Vehicles  | SC 627, IIT Bombay

Implemented multiple path planning algorithms using ROS both in (a) Gazebo (b) on hardware using a TurtleBot3 robot with access to LiDaR and odometry sensors. The algorithms were the (i) Bug-1 Algorithm (ii) Artificial Potential Function (APF) Algorithm (iii) Voronoi Cell Decomposition Algorithm (iv) Trapezoidal Cell Decomposition Algorithm.  

Multi-Agent Swarm Algorithms  | SC 627, IIT Bombay

Achieved (a) balanced and (b) synchronous consensus of a multi-agent network of TurtleBot3 robots functioning as a unicycle using the gradient control law.  

Flexible and Extendable 2D Manipulator | ME 423, IIT Bombay

Replicated a paper by researchers from Kyushu University to design a 2D manipulator robot based on the design of a woodpecker's tongue for cleaning of narrow pipes and ducts. Implemented the complete product pipeline, from need analysis to kinematic analysis and wrote the code for the automatic control of the mechanism.  

TossBot: A Robotic arm for executing Projectile Motion  | ME 604, IIT Bombay

Constructed a 3-degree of freedom robotic manipulator capable of tossing small objects. Interfaced servo motors to the robot arms via an Arduino for kinematic control. Worked on the analytical model for the robot using manipulator kinematics and rotation transformation matrix methods. 

Control of two-mass spring system | ME 311, IIT Bombay

Implemented controllers for ensuring stability of a mass spring system in MATLAB in the presence of (i) input saturation (ii) sensor white noise (iii) input disturbance. Designed controllers using different techniques (a) Linear Transfer Functional feedback (b) Full State Feedback. 

Sensor Modelling for Small Satellites | IIT Bombay Student Satellite Program

Studied modelling methods for (a) Gyroscope (b) Sun Sensor (c) Magnetometer (d) GPS  sensors for the Attitude Determination and Control of small satellites. Studied statistical signal processing concepts for analysis of noises (i) Bias instability (ii) Rate random walk (iii) Angle random walk in these sensors. 

Structural Analysis of an Antenna System | IIT Bombay Student Satellite Program

Performed extensive (a) harmonic (b) modal (c) random Finite Element simulations in Ansys Workbench for assessing the structural integrity of a deployable antenna system for the SANKET module. Supervised the complete manufacturing of the auxiliary module, right from vendor shortlisting and material procurement to final completion. 

Thermal Analysis of Satellite in Low Earth Orbit | IIT Bombay Student Satellite Program

Performed thermal analyses in Ansys Workbench for ensuring that all the electronic components in the satellite interior remain within optimal operating temperatures. Modelled the components in SolidWorks. Generated heat flux data using a Matlab script. Experimented with various passive thermal control methods (a) Heat Sinks (b) Thermal Pads (c) MLI (Multi-layer Insulation) and checked their efficacy. 

Graph Theory | Reading Project, IIT Bombay 

Studied discrete mathematical and linear algebraic properties of graphs such as isomorphism and graph eigenvalues. Analyzed various graph traversal algorithms such as Depth First Search, Breadth First Search, Dijkstra and A*.