IIT Mandi undergrad at the intersection of classical robotics and deep learning. I build systems that navigate, perceive, and act — on real hardware and in simulation.

Currently extending JdeRobot's multi-agent framework for drone simulations as a GSoC 2026 Contributor, and doing navigation + arm control for Team Deimos' custom-built Mars rover.

Zero-shot vision-language navigation on TurtleBot3 in Gazebo Harmonic. Robot navigates to natural-language or voice object goals in unseen maps — no pre-built map, no hand-coded waypoints. Gemini 2.5 Flash parses instructions into a goal stack; YOLOv8x-World + MobileSAM + CLIP perception; two-layer DualMap promotes objects seen across ≥3 keyframes to a global A* nav graph. Includes optional TD3 RL navigator and Next.js + rosbridge GUI.

Python  ROS2  Gazebo  Gemini 2.5  YOLOv8  CLIP  Nav2  SLAM  TD3

🥉 Bronze Medal — Inter IIT Tech Meet 14.0. Autonomous 4WD mecanum warehouse robot: navigates shelves, scans QR inventory tags, logs to cloud. Custom ros2_control C++ hardware interface; Cartographer SLAM; Nav2 with MPPI controller (DWB failed on mecanum); CLAHE→ZBar QR pipeline; optical-flow lift safety. Whole stack Docker-ported to a Pi 4 in 5 hours after a power fault destroyed the Pi 5 the night before the demo.

C++  Python  ROS2  Nav2  MPPI  Docker  Next.js  PostgreSQL

Benchmark of 8 deep learning architectures for synthetic CT generation from brain MRI, targeting MRI-only radiation treatment planning. Covers Mamba regression (TriPlane, TriAxial, U-Mamba, SegMamba), diffusion models (Hybrid Diffusion, Diffusion UMamba), and GANs (Pix2Pix, UNIT). Best image quality: TriPlane Mamba (SSIM 0.856, PSNR 25.79 dB). Best dosimetric fidelity: U-Mamba (Gamma 1%/1mm = 93.3%, clinical threshold passed).

Python  PyTorch  Mamba SSM  DDPM  GAN  Medical Imaging

Google Summer of Code 2026 @ JdeRobot / TheRoboticsClub. Extended JdeRobot's RoboticsApplicationManager (RAM) with multi-process / multi-agent support so N peer agents share one Play/Stop lifecycle via Linux process signals + Gazebo pause sync. Demonstrated as a two-drone pursuit-evasion exercise on ROS2 Humble + Gazebo Harmonic + Aerostack2, where the "cat" drone chases a pre-programmed "mouse" drone.

Python  ROS2  Gazebo  Aerostack2  GSoC 2026

Custom-built Mars rover designed and fabricated by Team Deimos, Jetson Nano onboard. Software: ROS2, Nav2, MoveIt2, FastSAM scene understanding, EKF multi-sensor fusion (GPS + IMU + wheel + visual odometry), STVL 3D costmaps, elevation mapping, ROSBridge web dashboard, Docker. Stack developed on a Clearpath Husky + Kinova 7-DOF arm dev rig before rover deployment.

C++  Python  ROS2  MoveIt2  EKF  Nav2  FastSAM  Docker

Real-time keyboard key detection powered by a custom-trained YOLO model.

Python  YOLO  Computer Vision

Experiments and research tooling exploring AI/ML ideas and pipelines.

Python  Machine Learning

An interactive JavaScript project blending logic and coding challenges.

JavaScript