AWS Physical AI Blog
Visual Asset Management System 2.6 Release
Sequel to the VAMS 2.0 Release post (20 June 2024).
AWS Physical AI and Prototyping and AI Customer Engineering (PACE) are excited to announce the release of the Visual Asset Management System (VAMS) 2.6, the most substantial update to the open-source accelerator solution since we introduced the platform’s modern foundation two years ago. Where VAMS 2.0 focused on making the platform secure, configurable, and deployable across the most demanding environments, VAMS 2.6 focuses on what organizations do on top of that foundation, defining and running processing at scale, searching assets by where they are in the world, and, for the first time, operating a live deployment through an AI agent. This release helps enterprises turn growing libraries of 3D data, CAD files, point clouds, and spatial computing assets into a system that not only stores and secures those assets, but actively works them.
The centerpiece of this release is a ground-up overhaul of the pipeline, workflow, and execution model. VAMS has always been the registry and orchestrator: you bring the compute, and VAMS gives it an identity, wires it into AWS Step Functions, runs it against versioned assets, and tracks the lineage. In 2.6, that orchestration layer is rebuilt end to end. Pipelines and workflows are now reusable, database-scoped definitions with typed execution configurations, and a single pipeline can serve an entire conversion matrix through templates and tag schemas instead of requiring one pipeline per output format. Execution becomes multi-file and multi-asset, with a global executions view, the ability to abort or re-run a job, and file-upload triggers that launch a workflow automatically. The overhaul also adds a new AWS Deadline Cloud pipeline type for render- and simulation-farm workloads.
A new set of Pipelines, Workflows, and Executions pages in the web interface brings a live execution board and a visual workflow builder to the whole experience, so teams can launch, monitor, filter, and re-run processing jobs directly from the UI.
Underpinning the release are two architectural moves that mature capabilities VAMS 2.0 first introduced. The API tier moves to an Amazon API Gateway REST backbone with a single centralized authorizer, consolidating the multi-factor and claims logic that previously lived in individual handlers. Search is modernized to Amazon OpenSearch 3.5 with next-generation Serverless collections that can scale to zero when idle, and it gains a marquee user-facing feature: geospatial search and a map view across both assets and files, so teams can find assets by location and see them rendered on a map.
The most significant new direction in 2.6 is that VAMS becomes agent-operable. A new VAMS agent skill and a VAMS Model Context Protocol (MCP) server let an AI host, such as Kiro or Claude Code, search, audit, cross-link, and run workflows on a live deployment through natural language. Both run on the user’s own credentials and are read-only by default, so an agent can never do anything the operator couldn’t; the two-tier access control system VAMS 2.0 introduced governs the agent exactly as it governs a person.
VAMS 2.6 introduces a range of other notable enhancements. The Physical AI catalog grows with a new NVIDIA Cosmos 3 world-foundation-model inference pipeline and a coordinate-transform pipeline that reprojects point clouds between coordinate reference systems, alongside new in-browser viewers: an embedded PlayCanvas SuperSplat editor that is now the default Gaussian-splat viewer, an open-source That Open Engine viewer for IFC building models, and a Physna viewer for 3D geometric search. Continuing the security and sovereignty focus that defined the 2.0 release, this version adds an AWS European Sovereign Cloud deployment template, Amazon Cognito OIDC federation, advanced AWS Identity and Access Management (IAM) role customization, a web application firewall that blocks by default, and comprehensive provenance tracking: permanent asset lifecycle history, per-version file change history, and outbound sync tracking. The command-line interface also matures with a new Amazon Simple Storage Service (Amazon S3)-style sync command group and bulk download support.
Video 1. VAMS 2.6 in action, semantic search across assets; the in-browser SuperSplat editor viewing an asset built from a Gaussian-splat capture; the That Open Engine BIM viewer exploring an IFC building model and toggling viewer features; geospatial map search locating assets by position; and an NVIDIA Cosmos text-to-world workflow generating.
The road from 2.0 to 2.6
VAMS 2.6 didn’t arrive in one leap. Over the two years since the 2.0 release, five feature releases steadily moved the platform from a secure foundation to the agent-operable system it is today, first hardening the data model, then the tooling around it, and finally the intelligence layered on top. Here’s the journey.
2.0 — The foundation
June 2024
- CDK overhaul into 10+ nested stacks with config.json feature switches
- New ABAC/RBAC permission system (Casbin), replacing Cognito groups
- Security for regulated environments: GovCloud, FIPS, WAF, KMS, VPC
- Asset tagging & linking; Image and Potree point-cloud viewers
2.1 — Pipelines take shape
November 2024
- Reworked pipeline model: execution types, input parameters, asset-metadata inputs
- New GenAI 3D metadata-labeling and 3D basic-conversion pipelines
- Use-case pipelines split toward their own modular structure
2.2 — Assets & files, done right
September 2025
- Assets and files formally separated; multi-file and empty assets
- Upload/download rewrite on presigned URLs (retires scoped-S3 access)
- External identity-provider (IDP) authentication support
- Community milestone: first major external customer contributions
2.3 — Tooling & viewers
November 2025
- New VAMS command-line interface for large-scale automation
- Plugin-based viewers: CesiumJS, Babylon.js, PlayCanvas, PDF, video, text
- Search & asset-management overhaul with advanced filtering
- Custom Lambda authorizers with IP restrictions; Gaussian Splat toolkit
2.4 — Metadata & integrations
January 2026
- Metadata & schema overhaul: multi-entity, bulk CSV, versioning
- Partner integrations: Veerum viewer, NVIDIA Isaac Sim, Garnet knowledge graph
- Auto-trigger workflows on file upload; CloudWatch audit logging
- API-only deployment option (no website)
2.5 — Modern web & Physical AI
April 2026
- Website & docs overhaul (Vite, AWS Amplify v6, dark mode default)
- NVIDIA Cosmos inference and GR00T fine-tuning Physical AI pipelines
- New Three.js 3D/CAD viewer and experimental Needle USD viewer
- API keys, Cognito user management, and rich asset versioning
2.6 — The platform runs itself
September 2026
- Ground-up pipeline / workflow / execution overhaul with reusable templates
- REST API backbone + OpenSearch 3.5 with scale-to-zero Serverless
- Geospatial search and map view across assets and files
- NVIDIA Cosmos 3 world-foundation-model inference pipeline
- New viewers: SuperSplat Gaussian-splat editor and IFC BIM (That Open Engine)
- AWS European Sovereign Cloud deployment template
- Agent-operable: VAMS agent skill + MCP server
Where VAMS goes next: a core surrounded by bolt-ons
With the 2.6 core stable, VAMS will enter a new phase. Rather than growing as a single monolith, the platform will expand through modular, open-source bolt-ons, optional integrations that let a customer target a specific industry, technology area, or Physical AI use case while sharing one governed asset backbone, one pipeline control plane, and one access model. The principle is deliberate: keep the core small and reliable, and let bolt-ons meet customers where their work actually happens, in the training loop, in the engineering cockpit, in the digital twin, and at the scene. VAMS already runs as enterprise-wide infrastructure across dozens of use cases in aerospace and defense, energy, and industrial manufacturing, and the roadmap ahead organizes upcoming investment into five themes:
- The Physical AI compute and pipeline backbone. Orchestrate the full train/fine-tune/synthesize/evaluate loop without forcing customers off their preferred compute, on-demand cloud workstations for tools like NVIDIA Omniverse, managed SageMaker HyperPod training scenarios, purpose-built NVIDIA world-foundation-model toolchains, and DAG-based downstream orchestration, all under VAMS’s execution control.
- Spatial intelligence, search, and visualization. Move from storing assets to reasoning over them: digital-twin knowledge graphs, a CAD-aware engineering cockpit, vector-based similarity search across 3D formats, LIDAR and point-cloud deep-labeling and segmentation and expanded large-scene 3D streaming.
- Ecosystem interoperability and content distribution. Reach the tools that consume the assets: first-class Unity and Unreal integrations and versioned asset packaging, deeper connectors for Omniverse/Nucleus/Isaac Sim and PLM/DAM systems, and on-premises caching with expanded global content delivery for heavy spatial data.
- Core platform hardening. Continued investment in the foundation: personal workspaces, cross-region environments, expanded in-browser viewers and file diffing, an asset-compliance system that enforces schema and metadata standards by construction, and an overhauled subscription and notification system.
- Agentic operations. Tying it together, agent-based operations that let users drive broad data operations, run workflows and pipelines, and invoke bolt-on integrations through natural language, turning the capabilities above into agent-assisted actions rather than manual steps.
The throughline is a small, dependable VAMS core surrounded by targeted, open-source bolt-ons: customers keep one governed backbone for their spatial and visual data while assembling exactly the capabilities their industry and Physical AI use cases demand. That is how VAMS scales from a solid 2.6 core into a platform for Physical AI and beyond.
Get started
Deploy VAMS 2.6 from the open-source repository on GitHub, and explore the documentation to configure it for your environment. Already running an earlier version? Because 2.6 requires a search reindex, re-pointing of any direct API integrations, and porting of custom pipeline code to the new execution model, follow the v2.5 → v2.6 upgrade guide end-to-end before you deploy. We’d love to hear what you build: open an issue or start a discussion in the repository to share feedback and feature requests.