UCLA Mobility Lab’s CivicTwin is one of the first round winners selected by the US Department of Transport’s ARPA-I Challenge. CivicTwin uses AI and generative video models to reconstruct transportation corridors from street-view data, aiming to test autonomous vehicles and ITS 10 times faster than traditional methods.

Each ARPA-I first round winner receives $20,000. The winners will go through to a second round to be held in 2026 to compete for prizes worth a total of $700,000.

According to UCLA, CivicTwin creates high-fidelity, physics-grounded, editable digital twin models of real streets and corridors, directly from street-view video and images for the purpose of simulating real-time urban transportation.

CivicTwin is designed to eliminate overhead dependencies such as non-scalable multi-view capture, especially drone/aerial, instead using generative video models and visual-geometric foundation models to reconstruct and fill unobserved regions with plausible, structured geometry. The result is a simulator-ready twin that preserves measured detail where data exist, fills gaps where they don’t, and scales across corridors and cities with dramatically ten times lower cost.

The UCLA Mobility Lab argues that current AVs, delivery robots, and automated traffic management require massive, safe, repeatable testing, particularly for rare, safety-critical scenarios that are hard to observe or stage on real roads. Today’s best simulators struggle with the authoring burden, physics realism, and city-scale coverage needed for safety cases and operations planning.

The Lab says that CivicTwin’s capability rests on five integrated elements: (1) object-aware reconstruction from street-view data guided by foundation-model segmentation; (2) generative inpainting to handle occlusions and sparse viewpoints while maintaining lane/sidewalk/curb topology and material priors; (3) physics-property inference (materials, contact, friction) with export to standard formats; (4) a GPU-parallel runtime with multimodal sensor simulation and behaviourally realistic agents; and (5) scenario editing for long-tail risk coverage with continuous updates as the network changes.

CivicTwin is purpose-built to accelerate ADS/ADAS (automated driving or advanced driver assist systems) and traffic management development, validation, and safety-case generation. The twin enables closed-loop simulation with real-world data, systematic rare-event generation (e.g., collisions/surrogates and perception robustness under weather/lighting). Secondary applications, such as micromobility and ground-support robotics, should benefit from the same tooling.