The pipeline is built around four coordinated stages, all running on NVIDIA Jetson T5000 hardware at the Edge.

• Real-time 3D scene reconstruction. A live 2D camera stream is captured and run through a TensorRT-optimized monocular depth-estimation model. The result is a metric depth map and a colorized XYZ+RGB point cloud, generated with low latency and no stereo hardware (Piccinelli et al., 2024). Because the reconstruction uses a single monocular sensor, the perception stack can reduce camera count, cabling, and mechanical complexity while generating depth and point-cloud data for downstream robotics applications.
• ai-powered spatial reasoning. The resulting point cloud is passed to an on-device model that converts raw geometry into structured scene understanding, including layout elements, object categories, and oriented 3D bounding boxes. The output is a semantic map, not just unlabeled points, showing what each object is, how large it is, and how it is oriented in space, so planning and navigation logic can use it directly.
• Intelligent object and layout recognition. The workflow identifies and locates objects within the reconstructed 3D space, building a layout of the environment that captures what is present, where it is, at what scale, and how it relates to walls, floors, and other objects in full 3D. Path planners, manipulation stacks, and SLAM systems can consume that layout directly, turning raw perception into a usable map of the workspace.
• Low-latency pipeline acceleration. GMSL camera streaming, NVIDIA zero-copy multimedia memory paths, TensorRT inference, and the ROS 2 layer work together to reduce unnecessary buffering and memory movement across the perception pipeline (NVIDIA, 2026). Keeping data on accelerated processing paths from ingestion through inference and publishing helps reduce latency for closed-loop robotics workloads.