Perception Machine Learning Engineer
Autonomous Robots
DEKA Research and Development Corporation

During my tenure in autonomous robotics, I contributed significantly to two key projects: the Delivery Robot - FedEx ROXO and the Security Robot, focusing on enhancing robot navigability and environmental adaptability.

For the Delivery Robot, I spearheaded terrain classification within images, optimizing a segmentation model for RGB images. Overcoming challenges in distinguishing elements like curbs, grass types, and surface irregularities was pivotal for robust robot navigation. Integrating depth information via stereo cameras mitigated model limitations and enhanced environmental feature recognition for adaptive navigation, addressing class imbalance issues with the Focal Loss function.

Navigating evolving project scopes and managing data redundancy and perspective variations were ongoing challenges. Refining depth perception algorithms to handle complexities arising from reflections and challenging environmental conditions was crucial.

Methodologically, I employed deterministic methods using depth data for analysis, incorporating de-projection, noise reduction, voxelization, and edge mapping. I also developed multi-task models leveraging depth and RGB data for boundary detection, segmentation, and depth hole filling.

Key achievements included developing a sensor fusion model utilizing lidar, radar, mono, and stereo cameras for a comprehensive global occupancy grid, and efficient pipelines for active machine learning achieving high accuracy in semantic segmentation and scene text recognition for number plate identification. Algorithms were formulated to fill depth holes in stereo camera data, enhancing algorithmic scene segmentation.

The role extensively involved executing ROS and C++ algorithms (RANSAC, PnP, ICP) for automation within the project, demonstrating adaptability to technological constraints and innovation in computer vision and robotics methodologies.

More information can be found here DEKA