NVIDIA routes Blender-to-OpenUSD scene prep through subagents with per-job acceptance criteria

Making one object grabbable is the example worth reading twice. NVIDIA's post says the job takes coordinated updates to the object's semantic label, its rigid-body configuration, and its collision geometry ^[11]. The post groups collision shapes with physics properties, and ovphysx is the library that authors and checks physics ^[7][15], so two of those three edits land on one tool surface and the third is scene metadata ^[20]. Astra, the model behind the coordinating Codex agent, connects those requirements across subagents and determines which checks are needed before the workflow proceeds ^[12][5].

Each subagent owns a specific job and its acceptance criteria ^[8]. That line is the part I would read first in any agent pipeline that writes to an asset a robot then trains against.

SimReady validation evaluates the resulting assets against a target simulation profile ^[7]. A pass therefore says the asset satisfies that profile, and only that profile. Of the seven prep questions the post lists ^[13][18] - labels, collision meshes, materials, sensors, clean USD export, whether the robot can perceive the target objects, and whether the scene will pass validation before time goes into debugging it in Isaac Sim or Isaac Lab - the last is the only one answered by a document you can point at. For the workflow to transfer to your scene, your simulation requirements have to be inside that profile.

That asset prep, ahead of the training loop, is where robotics agent work gets stuck is NVIDIA's own claim ^[3]. The post does not report timing, pass rates, or any other measurements for the workflow ^[21].

Adoption is four moving parts before any of this touches a scene. Codex, powered by OpenAI GPT-6 Astra, coordinates and interprets results ^[5]. Hermes subagents deployed through NVIDIA NemoClaw call the Omniverse Libraries ^[6], with NemoClaw supplying the reference architecture for building them ^[17]. The post also names OpenClaw and LangChain as alternative harnesses, configured with different Nemotron models for vision, reasoning, and tool use ^[16], and Claude Cowork by Anthropic as an alternative main agent ^[14].

Of the nine actions the post says the agent must be able to perform, seven are edits or checks on the scene and two decide what happens when a check fails ^[19]. Safe, mechanical issues can be fixed automatically ^[9]. Decisions that depend on developer intent, such as an uncertain semantic label or physical behavior, are escalated to a human with the relevant context and a proposed next step ^[10]. Which issues count as safe is set by whoever writes each subagent's acceptance criteria ^[8]. A reviewer should ask to see that list.