ChampionXperience is pleased to feature Pekka Varis, Dell Pro Precision Ambassador and NVIDIA Omniverse Ambassador, as he shares how he developed Hermes 3D Agent, a workflow that transforms a single concept image into a physically valid NVIDIA Isaac Sim scene, and what he learned about generative 3D for robotics.

A robot that will work in a workshop has to be trained and tested among racks, bins, cables and hand tools long before it meets the real ones. Every one of those objects needs a true size, a mass and a surface the robot can touch. Building such a scene by hand takes an artist weeks. Generative AI can draw the same workshop in seconds, but a picture is not a scene: generated meshes float, pass through each other and come out at the wrong scale.

Hermes 3D Agent is my answer to that gap. You give it one concept image of the space you have in mind. It builds every part in context, sizes it to real-world scale, places it where the picture puts it, and delivers an Isaac Sim scene in which physics has had the last word. Hermes is now in beta. Here is how it works, step by step, and the lessons that apply to anyone building simulation content with generative tools.

Left: the concept image. Right: the Isaac Sim scene Hermes built from it, with every part at real-world scale.

Walkthrough: one rack, from picture to physics

1. Read the picture before spending anything. Claude, the reasoning model inside Hermes, reads the concept image first: the rack type, every shelf and board, an inventory of items with counts, and hidden structure such as rows of bins behind the front row. It also flags what physics will refuse, like a shelf with no support or a bin on a slanted rail. All of this goes into a manifest, the contract that every later step is checked against.
2. Detect and locate.SAM 2.1 separates the 4K design into item masks, while DINOv2 identifies each object. Next, the system fits a camera to the frame edges. This camera maps each pixel to a ray in 3D space. As a result, Hermes can place each item on the shelf indicated by its pixels.
3. Library first, generation second. An approved asset is reused on every rack. Only missing items are generated with image-to-3D models, from tight crops of the design. Each new asset gets its meaning stored as data (front axis, real size, how it mounts), so placement never guesses.
4. Generators measure, Blender builds. Generated meshes are good at chunky, irregular objects and bad at thin metal, flat sheets and readable text. So Hermes uses a generated frame only as a ruler and rebuilds every profile, gusset, cap and shelf in Blender from a parts kit. Labels, keypad legends and LCD screens under cover glass are real geometry, not painted textures.

Labels, buttons and the LCD under its cover glass are real Blender geometry, so they stay sharp in close-ups.

5. Physics has the last word. Items are lowered to half a millimetre above their support and settled with PhysX in Isaac Sim. A Blender pass then pulls each item back toward its place in the design without breaking contact. Hermes simulates cables as rods using the Newton physics engine. For example, ten EV charging cables with Type 2 plugs settle into their wire baskets in 5.4 seconds. Their length changes by less than 1%.

EV charging cables with Type 2 plugs, simulated with Newton and settled in their wire baskets.

6. Checks run before a person looks. Every build passes structure, inventory, contact, colour and overlay gates and gets two scores: intent (does it match the picture?) and physical validity (does it stand up?). A build that fails goes back for another fix round; only builds that pass reach a person.

Three lessons for your own pipeline

• The design is intent; physics is law. Which items exist, how many and in what order come from the image. Pose, contact and scale come from simulation. When the concept art shows something impossible, the system preserves physical accuracy and records the reason for any deviation.
• Store meaning as data, not code. A front axis or mount type written once on an asset fixes orientation errors on every future scene.
• Turn every correction into a check. When a person flags a mistake, fix the step that caused it and add an automatic check, so a fixed problem stays fixed.

A finished Hermes rack in Isaac Sim: Blender-built frame, library tools and bins, simulated cables.

Why the workstation matters

All of this development ran on a Dell Pro Max T2 tower with an NVIDIA RTX PRO 6000 Blackwell. Its 96 GB of GPU memory keeps Isaac Sim, Blender, the detection models and the scene resident at once, with no swapping between tools. A second workstation with an RTX 6000 Ada builds racks in parallel from a shared job queue. Keeping everything local speeds up iteration. When I change a rule, I can rebuild the scene and test it in Isaac Sim within minutes.

The same toolchain on the keynote screen

With Persona AI, we brought humanoid welding robots into a shipyard digital twin using NVIDIA Omniverse. The robots were trained in Isaac Lab. We rendered the scene on the same Dell and NVIDIA hardware for the NVIDIA GTC 2026 keynote.

Watch the Persona AI video. Read Persona AI’s story, the Dell Technologies article, or the Cineshare case study.

Try Hermes

Workshops are just the beginning. Hermes 3D Agent can help you create many types of environments for robotics simulation and training. If your team builds simulation scenes, digital twins or robot-training environments, apply for the Hermes beta and send one concept image of the space you need.

• Author
• Recent Posts

Pekka Varis

Dell Pro Precision Ambassador, NVIDIA Omniverse specialist and Isaac Sim developer at Cineshare

Pekka Varis is the founder of Cineshare in Helsinki, specializing in real-time cinematic simulations, digital twins, and robotics visualization. He is a Dell Pro Precision Ambassador, NVIDIA Omniverse Ambassador, and Isaac Sim developer. Cineshare is also a member of NVIDIA Inception.

Latest posts by Pekka Varis (see all)

• From One Concept Image to a Sim-Ready Robot Workshop - 24 October 2026