🚀 Spatial AI Agents for Industrial Facilities from Falkor!

[Semih Asil]

A New Dimension to Maintenance Planning in Industrial Facilities​

Falkor (formerly Kongsberg Digital) is bringing together NVIDIA Omniverse and Microsoft Azure AI infrastructure to automate maintenance planning and inspection prioritization processes for industrial facilities. This integration aims to overcome complex operational challenges through its spatial reasoning capabilities.

Limitations of Traditional Approaches​

Operating complex industrial assets such as offshore oil and gas platforms, chemical refineries, and grids presents significant maintenance and inspection planning hurdles. Traditional software and generative AI platforms typically process tabular data, maintenance logs, and text-based piping and instrumentation diagrams in isolation. These systems lack the spatial context to evaluate real-world physical constraints such as crane access, component clearances, work crew access paths, and dynamic facility layouts. This leads to excessive manual effort by engineering teams to gather physical evidence and conduct site checks before executing hazardous maintenance tasks. Resolving these bottlenecks requires physics-based spatial intelligence that can connect physical asset geometry to live operational telemetry.

Maintenance Planning with Spatial Intelligence and Physical AI​

At NVIDIA GTC Berlin, Falkor introduced AI agents with spatial intelligence that can see, simulate, and reason about physical operating facilities. Developed in collaboration with NVIDIA and Microsoft, the system runs on the Falkor Surface interface within Microsoft Azure AI infrastructure. This system evaluates 3D physical parameters along with equipment histories, rather than being limited to static paperwork.

In a reference industrial facility test, in a scenario involving 4,847 pipe flanges, an agent analyzed spatial exposure, historical inspection records, live process temperatures, and concurrent work orders to identify 25 high-priority candidate flanges for engineering review. This provided engineers with traceable reasoning and visual evidence for each recommendation. In broader maintenance planning, agents simulate physical clearances and equipment access paths using interactive physics engines, allowing operators to detect layout conflicts and test hardware plans in simulation before mobilizing field crews.

Automated OpenUSD Modeling and Digital Twin Integration​

Industrial facilities rarely maintain pristine, unified computer-aided design (CAD) files; they often rely on legacy drawings and disparate operational databases where the as-built reality differs from historical engineering plans. To bridge this gap, Falkor employs an automated CAD Import Engine that converts native engineering files into simulation-ready Universal Scene Description (SimReady OpenUSD) assets embedded with NVIDIA Omniverse libraries. Geometric assets are resolved directly back into the facility's digital twin knowledge graph, allowing scenes to be dynamically reconstructed when underlying CAD models are updated.

The Falkor Platform integrates modular NVIDIA Omniverse core libraries, including ovstage for scene graph operations, ovrtx for RTX ray tracing, ovphysx for physics computations, and ovstream for low-latency browser streaming. These libraries are powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs hosted in the cloud on Azure. In cases where existing 3D CAD records are unavailable, AI agents powered by OpenAI's GPT-6 Astra reconstruct OpenUSD environments directly from 2D isometric drawings and P&IDs, registering them to as-built laser scan point clouds accessed via Microsoft Foundry.

Competition and Future Perspectives​

In the field of industrial facility digital twins and asset integrity management (AIM), the spatial AI benchmark focuses on data interoperability between ISO 15926 and ISO 14224 facility taxonomies and deterministic physics simulation performance. Standard enterprise asset management (EAM) platforms like SAP Plant Maintenance and IBM Maximo track maintenance records, mean time between failures (MTBF), and risk-based inspection (RBI) intervals, but lack native 3D spatial awareness. This requires maintenance planners to manually reconcile 2D drawings with separate 3D design models in tools like AVEVA Everything3D or Bentley OpenPlant.

Falkor's architecture uses OpenUSD as an extensible scene description layer, combining hierarchical tag-based facility data structures with geometric bounding boxes, collision meshes, and kinematic joint constraints. In the industrial metaverse and spatial computing sector, Falkor's platform conceptually competes with spatial intelligence initiatives from firms such as Siemens (integrating Teamcenter and NX with Omniverse), Bentley Systems (iTwin platform), and Dassault Systèmes (3DEXPERIENCE). Falkor deploys Omniverse libraries (such as ovphysx and ovrtx) encapsulated on Blackwell architecture server GPUs, enabling remote pixel streaming over WebRTC (via ovstream). Furthermore, integrating multi-agent reasoning via Model Context Protocol (MCP) servers allows advanced language models to deterministically query spatial APIs, isolating concurrent operation (SIMOPS) conflicts and safety exclusion zones without requiring manual geometric drawing.