I lead a team selected from diverse disciplines to develop an end-to-end warehouse management system, in which sensor-less mobile robots autonomously pick up packages from one location and deliver them to another. All robots are orchestrated by a centralized navigation and communication framework, and this solution earned a top-10 placement out of 9,000 teams nationwide at Flipkart Grid 3.0. 

This project highlights an end-to-end maze-solving system that combines a digital twin, computer vision, path-planning, and real-world robotic execution to enable a MyCobot Pro 600 to autonomously navigate a 4×4 maze. The work demonstrates seamless integration between MATLAB-based kinematic simulation and Python-driven control, culminating in both virtual verification and a physical demonstration of robust maze traversal. 

A fully decentralized swarm system that uses triangular-lattice formations and multi‐force interaction (cohesion, alignment, repulsion, obstacle avoidance) to achieve robust, coordinated navigation in complex environments.

In the MASt3R-SLAM system, reliable loop-closure detection is critical for long-term trajectory accuracy in visual SLAM. The original appearance-based keypoint matching relied on an L2 (Euclidean) similarity metric for comparing high-dimensional image descriptors. While L2 offers strong discriminative power, its computational cost on large descriptor sets can limit real-time performance. 

The network is built on a U-Net framework, using MobileNetV2 as its encoder backbone. It features a series of downsampling blocks, and the features from these blocks are fed via skip-connections into a self-attention module in the decoder. After the attention stage, the decoder applies upsampling steps, followed by max-pooling and convolutional layers.