Real-Time Digital Twins for Smart Cities

The operational problem

SLAM has always had a ceiling: it produces geometry without understanding. A point cloud knows there is an object at a coordinate; it does not know the object is a collapsed storm grate, a damaged sign, or a vehicle flagged in prior incident reports. Public-safety agencies need both — perimeter awareness around the cruiser, and an operational record of the city the cruiser drives through — and neither can be produced by shipping footage to a cloud analytics service.

What Enfuse built

• Roof-mounted 360° depth and camera capture with IMU-fused vehicle odometry
• Graph-based SLAM with two-stage loop closure for drift-free long patrol routes
• Voxel-downsampled point cloud maintained on-device across a full patrol shift
• Semantic scene understanding via an NVFP4-quantized 72B instruction model
• Structured JSON entity output that populates a queryable digital twin
• Infrastructure-defect cataloging — potholes, signage, lighting, hazards
• Natural-language scene query for officers and dispatch

Architecture

• See — 360° depth + RGB capture and 6-DoF pose estimation on the vehicle
• Understand — semantic scene reasoning over multi-frame context
• Locate — observations registered into a globally consistent city coordinate frame
• Predict — trajectory, dwell, and change detection across repeat passes
• Act — defect tickets into city systems; threat awareness to the officer and dispatch

Hardware and software

• NVIDIA Jetson Orin 64GB (onboard perception and SLAM)
• Orbbec Gemini 2XL 360° depth sensor, 200 Hz IMU
• Enfuse Qwen2.5-72B-Instruct-NVFP4 (W4A4, 32K context)
• NVIDIA Blackwell-class server (B200 / GB200) for semantic reasoning
• Graph SLAM backend, DBoW3 loop closure, TensorRT, vLLM

Deployment environment

Vehicle edge plus on-premises agency data center. Perception on Jetson in the cruiser; the 72B model and digital twin hosted inside the agency boundary. No video egress.

Results

• Frontier-tier reasoning (MMLU 83.68) at roughly a 3.4x smaller memory footprint
• Real-time semantic scene descriptions alongside continuous SLAM mapping
• A living city twin that updates on every patrol pass rather than on a survey cycle

Reference architecture. Benchmark figures are published model results; deployment figures are engineering design targets for this pattern, not measured results from a named client deployment.