Perception-Aware Onboard Autonomous Corridor Navigation and Dynamic Obstacle Avoidance on NVIDIA Jetson Orin Nano

Autonomous mobile robots operating in indoor corridors must perceive obstacles, plan safe trajectories, and execute motion commands under strict real-time constraints, yet embedded pipelines often evaluate perception and planning separately, although perception uncertainty and inference latency affect closed-loop behavior. This paper presents a perception-aware local navigation architecture for a differential-drive mobile robot running entirely onboard an NVIDIA Jetson Orin Nano. A lightweight deep vision detector, optimized through ONNX and TensorRT with FP16 quantization, is integrated into a ROS 2 navigation pipeline designed around Nav2 conventions. A perception confidence index adjusts the balance between goal progress and obstacle clearance, while measured inference latency limits translational speed when the perception pipeline slows down. Evaluation covers six corridor scenarios with static obstacles, pedestrian crossings, a narrow passage, clutter, and low light. Across the two dynamic-obstacle scenarios (50 runs per method), the perception-aware strategy achieves 92.0% success, compared with 68.0% for classical reactive navigation and 84.0% for deep perception without adaptive control. In the S3 ablation, TensorRT FP16 raises success from 72% to 88%, the confidence term to 92%, and latency-aware speed adaptation to 96%. The control loop runs at 18.6 ± 1.2 Hz, with 28.4 ms mean inference latency and 9.7 W mean power draw. Given the limited number of trials, these differences are reported as descriptive evidence rather than as proof of a general performance gain.