• Designed a semi-supervised domain adaptation pipeline to improve the 3D detection performance of agricultural robot harvesters in unseen environments

• Conducted ML experiments for hyperparameter tuning and model quantization on Xilinx FPGA

• Implemented a multi-object tracking-based car occupancy registration approach and demonstrated it on a field robot in a real parking lot

• Introduced a novel task of "Pose-aware Panoptic Segmentation" in urban self-driving environments to simultaneously predict 6D instance pose and panoptic segmentation of the scene

• Proposed a Panoptic-DeepLab inspired model architecture and a new evaluation metric; Augmented model performance with depth uncertainty estimation and multi-scale semantic feature fusion modules

• Polished Pytorch skills by implementing end-to-end model training, testing and hyper-parameter optimization

• Worked on a novel visual-servoing based navigation for a farm robot which achieved 94.5% coverage of a two-acre farm with less than 2.4 crops missed per row

• Contributed to the development of perception stack for an autonomous beach monitoring robot. Finetuned and deployed TensorRT optimized YOLO V3 object detection on a Jetson Xavier NX for real-time application