harsh tewari

Thesis: Sensitivity of LiDAR-only Human Pose Estimation to Point Cloud Sparsity and Occlusion • Built a live ROS 2 LiDAR pipeline (~50 ms per frame): PointNet segmentation, DBSCAN clustering, DAPT/LiveHPS pose estimation • GPU inference cut segmentation latency 34x (198 ms to 5.9 ms) • Reduced DAPT pose error by 33% (226 mm to 151 mm) by regenerating training labels and retraining PointNet • Benchmarked robustness to sparsity, occlusion, noise and beam loss across six human-robot scenarios

Collected robot training data for Vision-Language-Action (VLA) models on a humanoid robot. • Collected and annotated hundreds of multimodal pick-and-place teleoperation episodes using MANUS data gloves and an Xsens motion capture suit • Supported iterative data collection and model evaluation with ROS 2 based pipelines and the LeRobot framework • Built 3+ custom NVIDIA Isaac Sim assets for a client's robotics simulation integrated with VR teleoperation

Automated cobot positioning inside a radar test chamber for the RF test team. • Integrated a JAKA Zu 3 cobot into Keysight's OpenTAP environment for automated movement in a radar test chamber • Developed C# plugins to move the cobot to multiple target positions with no manual positioning • Implemented coordinate transformations and safety planes so the cobot never reached chamber walls or test devices • Worked in Scrum sprints covering testing, validation and documentation

Working student. Developed navigation and integration software for autonomous mobile robots (AGVs). • Developed an AGV navigation interface in C++ and ROS 2 with data transmission over TCP to Bosch Navigator • Tested and validated robot localization using Bosch ROKIT • Integrated a Shelly button over MQTT to trigger AGV pick-and-place operations • Built a React web GUI for monitoring and controlling the mobile robots

Working student. Embedded software development and validation for hydraulic pump control systems. • Developed an SPI F-RAM library for the Infineon XMC4000 microcontroller and worked with SPI/I2C protocols • Created RTOS projects and debugged hardware with a logic analyzer • Designed and ran CAN test cases in Vector CANoe to validate a hydraulic pump control system

Sensor electronics and CAN validation for the university's electric Formula Student race car. • Designed ADC and Hall-sensor boards for vehicle integration using KiCad and Embedded C (STM32) • Developed CANoe integration test cases for brake-pressure and temperature sensing boards, validating sensor accuracy and CAN communication