로봇주행가능성 정보 적용을 위한 코스트맵 프레임워크 경량화

초록

In real-world robotic deployments, robust and high-speed traversability mapping is essential for reliable navigation. This work presents a lightweight costmap framework that converts per-pixel traversability predictions from RGB images into a robot-centric 2D gridmap. The proposed pipeline time-synchronizes depth, traversability, and odometry, reconstructs a body-frame point cloud, attaches per-point traversability values via camera reprojection, and applies voxel downsampling to reduce redundancy and sensor noise before projecting the result onto a 2D costmap compatible with the ROS2 navigation stack. To reduce computation, we convert a PyTorch-based traversability network into an optimized TensorRT engine, enabling the framework to maintain real-time gridmap updates within the compute and memory budget of an embedded platform.