Neupeak Robotics

Real-Time Streaming and Data Infrastructure

• Designed and implemented a low-latency, Rust-based pipeline for real-time streaming, synchronization, and recording of multi-camera and multi-sensor data, leveraging NVIDIA hardware-accelerated video encoding, achieving a low memory footprint and approximately 5% CPU utilization on a Jetson Xavier.
• Developed a data generation pipeline using SAM3 video tracking to expand labeled datasets by tracking objects across frames, enabling dataset scaling with less manual annotation and accelerating model deployment in new environments.
• Enhanced the host-level monitoring system on Jetson Xavier to execute multi-layered action commands, resulting in a 60% reduction in CPU load.

7-Axis Manipulator

• Developed incremental enhancements to a 7-axis manipulator calibration system by integrating external tracking measurements and optimizing URDF kinematic parameters to improve end-effector pose accuracy.
• Implemented a time-parameterization algorithm for a 7-axis manipulator to enable smooth and reliable trajectory execution, computing parameterized trajectories in <= 10 ms under complex trajectory conditions and reducing overall trajectory execution time by 2-4 seconds.

Object Detection Model

• Integrated architectural enhancements into an existing multi-head object detection network by introducing an additional detection head, improving feature representation and enabling more consistent and accurate object localization and classification.
• Implemented a 3-branch semantic segmentation network for real-time object detection during navigation with clear mask boundaries.

Isaac Sim

• Integrated a custom mobile manipulator with a 7-axis manipulator in Isaac Sim with ROS communications using OmniGraph nodes.
• Designed and developed a custom simulation environment replicating greenhouse fruit farms, integrating deformable physics into fruit models to realistically emulate gripper-fruit interactions during robotic manipulation tasks.
• Implemented a pipeline using Isaac Sim tools to generate synthetic semantic segmentation datasets, significantly reducing reliance on manual labeling and decreasing overall data annotation time.
• Developed OmniGraph nodes for procedurally generating partial simulation environments, such as dynamic spawning of objects, and facilitating faster testing of robotic capabilities across diverse experimental setups.

End-to-End Navigation Model

• Developed an end-to-end vision-based navigation model using Transformers and diffusion policies, enabling mobile robot navigation from monocular camera input while supporting platform-agnostic deployment.
• Utilized a large open-source dataset to train the model on human navigational strategies, enhancing its ability to adapt to new environments and lowering the cost of collecting new training datasets.

Void Robotics

• Implemented a ROS2 micro-ros agent on Jetson Nano using an ARM-based Docker container. Integrated the BlynkEdgent C++ package into ESP32 for multi-WiFi connectivity and set up Docker-based VPN services to enable access to ROS2 systems across different networks.

Children’s National Research Center

• Led a team to design and develop a modular navigation system that enabled autonomous functionality for a wheelchair, reducing prototype cost by 52% compared to market alternatives.
• Developed a ROS stack utilizing RTabMap for visual SLAM using Intel RealSense stereo camera, enabling autonomous navigation in dense and narrow environments with a 70% success rate.
• Implemented YOLO for object detection and integrated Open3D for obstacle distance estimation, reducing collisions with dynamic obstacles and enhancing navigation safety.
• Deployed on a Jetson Nano using docker containerization, maintaining power consumption below 10W.

Bariflo Labs

• Developed a ROS-based automation system for a Water Aeration Device, reducing manual intervention.
• Designed and deployed a Random Forest regression model on the cloud to analyze real-time sensor data with 160ms latency, ensuring 89% accuracy in aeration control predictions
• Integrated an ESP32-based Wi-Fi controller with a mobile application using MQTT and RESTful APIs, enabling remote monitoring and real-time control, reducing on-site maintenance costs by 40%.
• Optimized edge-to-cloud communication, reducing data transmission latency and ensuring 96% uptime for continuous aeration system monitoring.

National Institute of Ocean Technology

• Designed and fabricated a soft robotic gripper using flexible 3D-printed materials, enabling secure handling of fragile objects up to 500g.
• Developed and simulated the complete mechanism in SolidWorks, optimizing structural integrity through stress analysis.
• Developed mechanical linkages for underwater manipulation and integrated an ESP32-controlled stepper motor, ensuring precise actuation with a ±0.5mm tolerance.
• Optimized the gripper’s design, reducing overall weight by 30% while maintaining high grip strength, enhancing efficiency for underwater applications.

University of Maryland, College Park, MD, USA

Robotics Engineering

Kongu Engineering College, TN, India

Mechatronics Engineering