Scenario

An autonomous quadruped — Panza Robotics' Artaban, carrying Sensifai's on-device AI perception stack — patrols Castle XVII and inspects a scheduled set of locations across walls, ceilings & vaults, floors, and building components. At each location it scans with a multimodal sensor payload (LiDAR / RGB / thermal / acoustic) and issues a restoration triage verdict — stable, monitor, or restoration required — with defect evidence and a recommended intervention. All inference runs on-device; no raw data leaves the site, and every verdict is sealed to an audit ledger.

Grounded in the Call (CR.EU.IN. HERITAGE — SMP-COSME-2024-CLUSTER)

• Strand 1 — Product Innovation: a new-to-the-firm product (autonomous built-heritage inspection) reducing dependency on non-EU robotics and cloud platforms.
• CHVC domain: Safeguarding & Management of tangible cultural heritage.
• EU-based technologies (Option A): Artaban (SK) + Sensifai on-device AI (BE) + Pactus blockchain audit layer — EU-owned, EEA data sovereignty, EU AI Act compliant.
• Cross-border, cross-sectoral consortium: Sensifai BV (Belgium, Digital/AI) × Panza Robotics (Slovakia, Advanced Manufacturing).

Grounded in the Research (Three.js 3D Environment Architecture)

• Lightweight, dependency-free runtime: vendored Three.js and locally embedded GLB models — no external assets, near-instant time-to-interactive.
• Navigation: a baked navigation grid — the castle's own triangles are rasterised to detect walls and towers, then A* pathfinding with string-pulling gives the robot a Unity-NavMeshAgent-style route: given an origin and a destination, the quadruped finds and follows the path itself, steering smoothly, slowing into turns and braking at the stop.
• Unified camera state machine: a single perspective camera tweening between free-orbit and follow modes via lerp/slerp — no camera recreation.
• Cheap lighting & pooled math: dusk-style lighting with a single shadowed sun, and pre-allocated vectors/quaternions (zero per-frame allocation) for a steady 60 FPS.

3D model credits

“Castle XVII” by gogiart (@agt14032013) — Sketchfab, licensed under Creative Commons Attribution (CC BY). Textures optimised (≤1024 px, JPEG/PNG) and the GLB embedded locally for the demo.

“Robot dog Unitree Go1” by C. Yamahata (@cyamahat) — Sketchfab, licensed under Creative Commons Attribution (CC BY). The rigged GLB is embedded locally and animated with its own walk (“1LYP”) and idle (“1Idle”) clips.

Production deployments would use the research's full stack — WebGPURenderer, three-mesh-bvh kinematic capsule, navcat pathfinding, Meshopt + KTX2 assets.