KINBOT-AI: Design, Implementation and Controlled Demonstration of a Resource-Constrained Edge-AI Mobile Robot for Mixed-Traffic Observation in Kinshasa

Mixed traffic in Kinshasa combines cars, motorcycles, and pedestrians on roads with limited fixed sensing and intermittent connectivity. This paper presents KINBOT-AI as a design, implementation, and controlled-demonstration artifact: a resource-constrained mobile Edge-AI platform that performs local perception, tracking, counting, queue estimation, safety control, event logging, and human-supervised alert handling. The evaluation comprised 40 navigation missions on a closed 85 m route, 25 counting sequences, 15 queue scenarios, 12 auxiliary environmental traces, 20 simulated alerts, and eight forced 30 s network outages. The prototype completed 37/40 nominal missions (92.5%; 95% Wilson interval: 80.1–97.4). Classwise F1 scores were 0.87 for vehicles, 0.82 for motorcycles, and 0.79 for pedestrians. Counting errors recalculated from the reported manual and automatic totals were 8.6%, 14.1%, and 16.5%, respectively. Queuelength MAE was 1.7 units, average autonomy to 20% battery was 58 min, and median alert latency was 1.4 s. During all eight outages, local obstacle avoidance continued and no logged event was lost or duplicated after reconnection. These measurements characterize the prototype only within the controlled study environment and must not be interpreted as citywide operational performance.