AI can generate a 3D model in minutes. Turning that model into an asset a robotics team can trust, reuse and find again next month is a different job. That job is what we built Hermes 3D for.
Hermes 3D is an AI asset production agent developed at Cineshare. It takes a concept image and works it through to segmented, inspected 3D assets in OpenUSD, reviewed in NVIDIA Isaac Sim. A person approves every result. Hermes is now in private beta, and we are looking for a small group of teams to test it on their own asset ideas.
The problem: one-off generations do not make a library
Simulation scenes need a lot of believable content: racks, shelving, bins, tools, foam organisers, cables. Robotics teams training in Isaac Sim need many of these, in variations, and they need them to behave like real objects in a USD scene.
Generative 3D tools can produce single models quickly. They rarely tell you where a model came from or which prompt made it, and they don’t help you make a second version that still fits the first. You end up paying to generate the same bin twelve times, each one slightly different, none of them traceable. Hermes treats asset production as a pipeline with a memory instead.
One real case: the workshop rack
Our first full test case is a workshop rack. From a single concept, Hermes identifies the structure and what it holds: the frame, the tools, the foam organisers, the bins and their contents. It then works on each asset family separately. It prepares clean source images, generates segmented meshes, inspects the geometry and places the results in Isaac Sim for review.
1 · Empty structure
2 · Filled with contextHermes knows the full context of what it is building. A screwdriver belongs in a foam organiser, which belongs in a drawer, which belongs to this rack. That context lets it make reasonable variants, such as a second rack with a different tool set, without breaking the logic of the scene.
How Hermes works, step by step
- Concept and image preparation. Image models produce concepts and source-specific images, including complete empty structures, foam organisers, tools and bins.
- Visual understanding. A vision model reads each source image, classifies the asset family and flags uncertainty or cropping. Weak inputs are caught here, before any 3D generation is paid for.
- Segmented 3D generation. Hermes generates detailed meshes split into meaningful parts. It records the exact source image, the settings, the task IDs and the resulting files.
- Smart duplication. Twenty identical bins do not need twenty generations. Hermes duplicates bins and other repeated assets instead of generating them again, which saves time and cost and keeps the scene consistent.
- Geometry inspection. Local scripts and Blender check segments, dimensions, transforms, intersections and which component owns which part. Qualified candidates are converted from GLB to USD.
- Omniverse materials and assets. Hermes uses native NVIDIA Omniverse library materials and assets, so the results look and behave like the rest of an Omniverse scene.
- Review in Isaac Sim. Isaac Sim renders multiple views and supports placement, alignment, material exploration and bounded physics experiments.
- Production memory. Structured records keep asset identities, prompts, hashes, versions, costs, render evidence and every human decision.
Under the hood, Hermes is Python orchestration with JSON manifests, GLB and USD assets, and tool connections that link each stage. Every component keeps its identity and history as it moves through the pipeline. That makes results easy to review, reproduce and reuse.
Why a person approves every result
Industrial scenes are full of parts that must look right, and an agent that guesses is a liability. In Hermes, AI models provide analysis and recommendations, and final acceptance is always a human decision. Your reviewer accepts an asset, rejects it or sends it back. Approved assets are preserved as the project develops, so good work is never regenerated by accident.
We are also clear about the limits. Hermes assets are reviewable scene candidates, not mechanically validated parts. Your engineers and simulation tools remain the source of truth.
What comes next
- Realistic cables. Cables in Hermes scenes will be simulated with Newton physics in Isaac Sim, so they hang, bend and move like real cables instead of sitting as rigid meshes.
- SimReady conversion. We are adding conversion to SimReady assets, so Hermes output carries the structure and properties simulation workflows expect.
- More autonomy, same control. Generation and inspection workflows already run. We are moving Hermes toward greater autonomy step by step, with human approval kept at the centre.
Who Hermes is for
Hermes is built for teams that need many believable 3D assets and need to know where each one came from. That includes robotics and physical AI teams building asset libraries for Isaac Sim, and industrial and digital twin teams with workshops, racks and equipment to model. It also suits manufacturers with product images but no clean 3D, and studios that need 3D props for video and product storytelling.
Join the Hermes 3D beta
Hermes is not on sale yet. We are inviting a small group of teams to test it on their own asset ideas. Beta testers get a small set of 3D assets built from their concept or reference images, converted to USD and reviewed in Isaac Sim, with a direct line to the people building Hermes. Joining requires no purchase or commitment.
