Sumukh Balakrishna

1.Engineered a real-time in-cabin monitoring system for driver presence detection, optimizing YOLOv8 models via TensorRT FP16 quantization to achieve 94% mAP with 25ms latency on NVIDIA AGX Orin. 2.Established accuracy-latency benchmarks for semantic segmentation by training and evaluating SOTA architectures (SegFormer, DeepLabv3+, HRNet); selected and deployed EfficientVIT for drivable area segmentation, achieving 85% mIoU at 45ms on NVIDIA Jetson via pruning and FP16 quantization. 3.Implemented a Transformer-based object detection pipeline using RT-DETR to identify generic obstacles, optimizing the model for real-time performance to strictly meet <50ms latency requirements on the autonomy stack. 4.Led the end-to-end development of traffic light recognition systems, managing the full lifecycle from site-specific data collection to deployment and integration within RTMAPS middleware on Jetson Orin. 5.Spearheaded the computer vision component for an aircraft wing tracking POC, training and validating custom detection models on site-specific datasets to enable autonomous operations in complex airport tarmac environments.

1.Designed and implemented a vision-based grasping system for unknown objects utilizing Point Cloud Library (PCL) for plane segmentation, Euclidean cluster extraction, and filtering. 2.Utilized a Concave hull to estimate centroid and grasp points ensuring secure grasping of objects without collision 3.Increased 2D and 3D sensing capabilities by avoiding perspective projection and outliers due to sensor noise 4.integrated analytical grasping algorithm with learning based and Incorporated Moveit! for pick and place 5.Conducted rigorous testing on grasping and manipulating objects from YCB object dataset in both Franka Emika Panda real robot and Gazebo to validate the effectiveness of benchmarking pipeline 6.Collaborated with developers at Umass Lowell Nerve Center for testing algorithms on their experimental setup and derived comparison for benchmarking various vision based grasping algorithms