Repository navigation
Releases
USD Content Agents v0.6.2
Content Agents 0.6.2 (20 Oct 2026)
Native Windows execution is unsupported in the 0.6 release. On a Windows
host, run the supported workflow inside WSL2.
Changed
- The default hosted model for VLM and LLM calls is now
nvidia/nemotron-3.5-super-vl-120b-a12b. It replacesmoonshotai/kimi-k3,
which hosted NVIDIA NIM has retired.
Fixed
- The Material Agent REST service honors
layer_onlyoutput. Pipeline jobs
keep the material layer unflattened, and
/artifacts/{session_id}/output-applieddownloads the applied layer instead
of the flattened render copy. - usd-cli OVRTX rendering no longer produces black images when
CUDA_VISIBLE_DEVICESis set or when a scene relies on the default dome
light. The optionalrender.ovrtx_active_cuda_gpussetting pins rendering
to specific GPUs explicitly. - Hosted NVIDIA NIM requests that hang or fail with HTTP 500, 502 or 504 are
retried. With the default timeout, each attempt reads at most
WU_NIM_ATTEMPT_TIMEOUT_SECONDS(default 180 s); an explicittimeout
applies to every attempt as is. A hung request gets up to three attempts,
and hosted NIM clients default to a 600-second request timeout. A request the
endpoint has accepted with HTTP 202 is never sent twice. - Remote-render sensor buffers with a non-square render size, such as a
640x480 depth buffer, are reshaped to the requested size instead of a
guessed square, which produced a misshaped depth image. - Raised aiohttp to 3.14.5 in the locked installs (constraint floor 3.14.4)
to clear CVE-2026-111064, CVE-2026-111063 and GHSA-2g87-pp6c-29x5. - Raised Pillow to 13.0.0 to clear BDSA-2026-37854, BDSA-2026-37841 and
BDSA-2026-37860.
Full history: see CHANGELOG.md.
USD Content Agents v0.6.1
Content Agents 0.6.1 (15 Oct 2026)
Content Agents 0.6.1 is a maintenance release for the 0.6 line. It fixes hosted
NVIDIA NIM vision-model workflows, adds a validated material-layer output and
an explicit visual-review opt-out, and completes the bundled-asset notices.
Native Windows execution is unsupported in the 0.6 release. On a Windows
host, run the supported workflow inside WSL2.
Highlights: hosted NVIDIA NIM fixes
- Configured endpoint and image limits are respected. Requests now use
the hosted NIM endpoint you configure, instead of a model-specific route
that could be rejected. Each model's image limit is applied automatically.
For example, multi-view material classification with hosted
meta/llama-3.2-11b-vision-instructcombines the rendered views into one
numbered grid image rather than failing with HTTP 400. - Hosted VLM rate limiting no longer stops the workflow. Concurrency to
hosted NIM VLMs is capped, and HTTP 429/503 responses are retried with
jittered backoff instead of a burst of immediate retries. Shared request
deadlines and explicit provider timeouts prevent unbounded waits. A model
endpoint check reports misconfiguration before a run starts.
Added
- Material assignment can publish a validated material layer over the source
USD, preserving the source geometry and composition. - Select an installed Codex executable with
CONTENT_WORKFLOW_CODEX_EXECUTABLE. The pinned executable remains the default;
invalid paths fail without fallback. - Explicit
--skip-vqafor Material assignment and Physics apply. Authoring
checks remain required, Physics still requires passing real simulation, and
skipped visual review is reported as unverified with conditional evidence.
Physics tuning and refinement do not support this opt-out.
Fixed
- Improved hosted NVIDIA NIM routing, model endpoint checks, image-limit
handling, concurrency limits, and retries for throttled or unavailable
endpoints. Shared request deadlines and explicit provider timeouts prevent
unbounded waits. - Prevented command locks from remaining owned by a long-lived launcher when
an outer action group cannot be identified. - Preserved delegated visual-evidence integrity when Material visual review is
explicitly skipped; skipped Physics reports no longer imply rendered review. - Strengthened Git LFS payload verification during distribution and corrected
the recorded 0.6.0 release date.
Changed
- Removed sample assets and documentation media that no example or test used.
The Material ladder reference images are now renders of the bundled ladder
model, and the usd-cli Cube sample is generated by a bundled script. ASSET_NOTICES.mdlists every bundled asset file with its source and
terms; a rootNOTICE,THIRD_PARTY_NOTICE.mdand the README point to it.
See CHANGELOG.md for the complete history.
Astra asset experiment: submitted data companion
This experiment-data companion supplies exact submitted USD assets, bindings and required dependencies for five cases in the Astra Ultra comparison.
The archive contains 26 unchanged payload files (212,177,755 bytes), with 45 ordinary files including notices, review records and a verifier. Both arms of refrigerator, gripper, Thor robot arm, excavator and Dextra hand remain represented. The Content Agents refrigerator has no submission; its gripper artifact is incomplete because bindings are absent. No failed artifact was repaired for publication.
Archive: additional_publication_v2.tar.gz (95,306,294 bytes)
SHA256: a58fe0ba18b75f1917a17d5dfa12655e4d7d24ef035a3aab2b2ba20a203d6e30
Contained manifest SHA256: 7e6b0d27b703899a92a3df5aea6613a1d3fe6247c1f5543373e803e7880898cf
After checking the digest, extract into a fresh directory and run python3 -B verify.py . from the extracted additional_publication_v2 directory. Optional decoded USD verification requires usd-core==25.5. No author program, model or simulator is executed by the verifier.
The archive retains attribution, original notices, modification notices and CC BY / MIT / CC BY-SA terms for the corresponding assets. Native metadata discrepancies and vendor-provenance caveats remain documented. Cases 05/06/09 are withheld under the recorded provenance/source limits. Original source/reference closures and simulator runtimes for these cases are not supplied, so this is not a complete ten-case replay.
This prerelease distributes experiment data independently of the project software release series. See the comparison report for independent task outcomes and the exact verification boundary. The delivery receipt records the earlier local packaging step; its historical not-uploaded status is preserved.
USD Content Agents v0.6.0
Content Agents 0.6.0 (07 Sep 2026)
Content Agents 0.6 makes the adaptive, evidence-driven Content Workflow the
default public entry point for supported asset tasks. It integrates governed
scene operations through usd-cli and strengthens resumability, provenance,
packaging, and release safeguards across the public stack.
Agentic workflow runtime targets are native Linux and WSL2. Native Windows
execution is unsupported in the 0.6 release. On a Windows host, run the
supported workflow inside WSL2. Agentic WSL2 workflows use remote OVRTX because
the local Vulkan renderer is unavailable there. Fixed pipelines use the Warp
rendering backend under WSL2 only.
Added
- Added root-discoverable Content Workflow skills for composed asset,
articulation, material, texture, physics, validation, Geometry, repair,
SimReady, large-scene, and mesh-segmentation work. Durable runs retain typed
handoffs, review decisions, reusable memory, exact artifact identities, and
interruption-safe resume evidence under rootruns/. - Integrated usd-cli as the governed low-level scene-operation layer for
inspection, conversion, editing, rendering, material binding, physics,
validation, and live workflow viewing. Local and remote OVRTX execution,
package-owned telemetry, and workflow-owned acceptance evidence remain
distinct and fail closed when the selected renderer is not ready. - Retained the Joint Research Preview's established revolute/prismatic boundary
while moving Joint, Texture, and Validation execution into the durable
long-running coordination layer with explicit custody, truthful recovery,
and resource teardown. - Added rendered Material refinement and variation authoring and agentic
Physics refinement evidence. - Added opt-in, provenance-bound VoMP mass-property authoring to agentic Physics
and prompt-first, material-aware Texture weathering with bounded quality
checks. - Added provider-free Texture UV preparation and selective-leaf execution for
scoped application, evidence, and publication. Composed workflows retain
exact leaf, dependency, artifact, and terminal-receipt identities and fail
closed when UV, provider-result, or custody requirements are not met. - Added complete composed-asset catalog coverage for Material, Physics,
SimReady conformance and validation, portable USDZ packaging, independently
bound Articulation and final-package OVRTX evidence, and one exact combined
terminal result. Requests can require capabilities, dependency edges, and the
sole terminal up front, so an incomplete installation or graph stops before
model, renderer, authoring, or runtime work instead of beginning a partial
run. - Added provider-neutral Geometry workflows, self-contained Geometry Repair,
SDF Tools, and OpenVDB Runtime packages, source-free semantic CAD handoff,
dependency-closed non-USD intake, and exact render-bound mesh-segmentation
evidence. Native B-rep authoring remains an optional, separately authorized
deployment capability rather than a public package.
Changed
- Replaced the public NVIDIA NIM chat and VLM defaults with
moonshotai/kimi-k3across shared model tooling, Material, Physics, Joint,
Texture, Validation, services, benchmarks, deployments, and examples.
Kimi-K3 reasoning effort is model-aware and defaults tomax. Texture image
generation remains on its separate image model; Kimi-K3 supplies its
text-only auto-prompt path. - Moved Onshape authoring to the official Onshape Labs FeatureScript MCP and
added an optional local handoff for immutable STEP, glTF, and OBJ exports.
Onshape credentials remain local and are not accepted by the Geometry Agent
service or Helm deployment. - Moved the unreleased Geometry Authoring implementation and its direct-OCP
dependency behind explicit internal boundaries. Public staging, packages,
images, locks, notices, and release workflows no longer ship that stack. - Start setup, authentication, workflow commands, and coding-agent sessions
from the repository root. Release-approved skills are discoverable from the
root;agentic/remains an implementation and package reference rather than
a required working directory. - Existing
material-agent,physics-agent,joint-agent,texture-agent,
andvalidation-agentCLIs, YAML configs, Python APIs, benchmarks, and REST
services remain available as explicit fixed-pipeline interfaces. The
$fixed-pipelineumbrella keeps their detailed guidance separate from
adaptive workflow routing. - Replaced the retired Content Workbench scene transport with workflow-owned
usd-cli sessions and renamed scene-operation timeout options accordingly. - Updated OVRTX to
0.4.1.364340, with version-bound isolated-runtime readiness,
complete-lock-digest readiness, verified side-by-side replacement and
rollback, updated Python 3.11/3.12 and x86_64/ARM64 runtime locks, and aligned
rendering-service documentation and images. WSL2 cannot run local OVRTX
because it lacks the required Vulkan device; use remote OVRTX for path-traced
or material evidence, while supported mesh-only contexts may use the
CUDA-onlywarprenderer. - Reworked public READMEs, skill routing, package inventories, native builds,
service contracts, and dependency notices so public copies, wheels, and
source distributions stay aligned with the release source.
Fixed
- Setup now checks for
unzipbefore fetching build resources, documents the
resource-skip path, and accurately distinguishes the fixed-pipeline CLIs
installed by general setup from the separately installed Validation CLI. - WSL2 physics runtime routing recognizes the
/dev/dxgGPU bridge. Remote
rendering can authorize a declared run artifact root so generated sibling
textures remain portable without weakening traversal and symlink confinement. - Geometry deployments can explicitly opt into an unauthenticated identity for
an operator-controlled remote-render sidecar; authenticated endpoints remain
the default. The Helm chart package version reflects that template change. - Native-Windows development clone guidance enables Git long paths while the
supported 0.6 execution route remains Linux or WSL2. - Preserved variant selections in OVRTX asset-opinion identities so textures
and MDL modules authored directly on a variant-owning prim cannot overwrite a
sibling variant during export. Remote package-member texture arrays remain
untouched when local USDZ textures are localized for rendering. - Localized complete MDL and texture dependency packages at the shared OVRTX
export boundary for local and remote rendering. USDZ-backed materials now
preserve MDL imports and resources across variants and instances without
mutating the source package, while unsafe recursive directory layouts and
incomplete rewrites fail closed for every workflow using OVRTX. - Hardened public service images with digest-pinned bases, current distro
packages, and an image-only media contract. Public Geometry and Physics no
longer encode or decode video or install FFmpeg, imageio-ffmpeg, or OpenCV;
they publish PNG evidence frames, with Pillow GIF previews where applicable.
Material keeps its separately owned codec-free OpenCV build. The public
dependency closure also dropspyte; video and terminal-replay demo skills
and their dependencies are internal-only. The public Geometry 0.6 package
and image also no longer ship the CAD converter or its unqualified embedded
Python runtime. CAD and mesh sources remain immutable handoffs and fail closed
before conversion unless a separately qualified converter or provider-produced
USD is available. - Prevented public staging from copying NVIDIA-managed Step1X runtime
definitions, dependency locks, and component/model inventories. The
operator-managed Texture Variation API adapter remains available. - Strengthened Geometry source admission with private dependency snapshots,
undeclared-UDIM rejection, and retained provider units and axes. Provider
provenance inputs now remain digest-bound across later revision and export
requests. - Preserved Geometry provider-bundle dependency closure and source coordinate
metadata through normalization, semantic OVRTX view planning, connector
handoffs, and service publication. Manifest-only B-rep inputs now repair the
admitted selected representation before USD conversion, while accepted repair
derivatives remain bound to their worker-verified digest through consumption. - Made the standalone OVRTX rendering service compatible with the usd-cli
remote protocol used by Content Workflow. Remote readiness, bounded USDZ
uploads, complete frame and camera coverage, renderer-reported execution
metadata, and one-time daemon recovery now fail closed together while the
legacy rendering endpoint remains available. Ordered, repeated, and
fractional USD time codes now retain their exact response association. - Fixed writable USDZ rendering so managed camera edits and package-relative
textures reach OVRTX together without modifying the source package. Untouched
packages remain byte-reused, while edited packages retain valid render-product
camera targets instead of producing zero render products. - Fixed strict usd-cli verification so dependencies stored inside the inspected
USDZ package remain package-local instead of being reported as external
filesystem dependencies. - Fixed Texture UV verification for self-contained USDZ inputs so deterministic
projection hashes stable authored asset operands instead of transient package
extraction paths while retaining exact source and dependency byte bindings. - Fixed Texture discovery, provided-image authoring, and scope validation for
uniquely authorable internal instance proxies while retaining fail-closed
handling for external or multiply instanced references. - Fixed skill-routed Texture resume after interruption at the provider-free
usd-cli...
USD Content Agents v0.5.2
Content Agents 0.5.2 is a focused public hotfix over v0.5.0.
Changed
- Restored
google/gemma-4-31b-itas the public NIM default for VLM and LLM workloads across Material, Physics, Joint, Texture, and Validation agents. - Aligned shared model helpers, services, collection deployment, examples, public documentation, UI metadata, and regression coverage with the restored default.
- Updated public clone, setup, and agent-guide references to the renamed
NVIDIA-Omniverse/usd-content-agentsrepository. - Preserved explicitly Qwen-specific self-hosting workflows and benchmark configurations as opt-in model-specific paths.
Full Changelog: v0.5.0...v0.5.2
This release was re-cut on 31 Jul 2026 to retire meta/llama-4-maverick-17b-128e-instruct, which reached end of life on 27 Jul 2026 and returns 410 Gone, along with the meta/llama-4-maverick-17b reference that returns 404. Both are replaced with google/gemma-4-31b-it, and the retired model is no longer offered in the Material Agent service VLM list.
The active model defaults were already google/gemma-4-31b-it, so this affected an optional selection, a disabled-by-default evaluation judge, a generated CLI config example, and docstrings.
USD Content Agents v0.5.0
Content Agents 0.5 adds the new Joint Agent Research Preview and expands the
suite with asset-specific material creation, REST-accessible Physics
refinement, reference-conditioned Texture editing, stronger Validation, and
adaptive workflow review and repair.
At a glance
Material Agent — Beta
- Creates new, asset-specific textured materials from reference images and
optional text guidance, then assigns them through a temporary USD library. - Adds selectable SimReady catalogs with 265 curated or 1,529 currently usable
materials, plus clearer per-part assignment evidence. - Generated-material mode is single-asset and does not include full native
OpenPBR/MaterialX authoring.
Physics Agent — Beta
- Exposes iterative simulation refinement through REST: run trials, compare
against text or visual goals, revise, and repeat. - Returns the tuned USD, a recording when produced, and evidence, with improved
friction handling, failure reporting, and ARM64 support.
Joint Agent — Research Preview
- Adds revolute joints for rotating parts and prismatic joints for sliding
parts, including accepted connections, axes, and limits. Aggregate or
multi-root assets also receive exact link membership and articulation roots. - All 17 qualification packages passed the selected static SimReady checks,
covering 78 joints and 119/119 Foundation results. - Adds articulation structure but does not create missing rigid bodies, masses,
colliders, or drives, or prove dynamic behavior.
Texture Agent — Research Preview
- Adds reference-conditioned, UV-aware editing for selected materials or parts
through an operator-provided Texture Variation service. - Supports bounded planning and targeted retries while preserving accepted
textures and modeled surface detail. - The default remains text-to-texture generation; a managed Step1X runtime is
not included.
Validation Agent — Research Preview
- Adds direct local OVRTX rendering and more consistent, fail-closed execution
across supported render and model providers. - Continues to validate rendering, visual appearance, authored physics, and
behavior evidence through CLI and Python.
Agentic Workflow — Research Preview
- Uses a persistent Workbench session to inspect intermediate renders, correct
fine part-level or cross-view inconsistencies, and review Material results
again, with up to three passes by default. - Supports durable decomposition, per-asset processing, collection, and
recovery for larger Material scenes. - Does not yet provide agentic Texture or Joint workflows or orchestration
across all five agents.
Agent details
Material Agent — Beta
Material Agent can now create new, asset-specific textured materials from
reference images, with optional text guidance. It generates the required
texture maps, packages them into a temporary USD material library, and uses the
new materials during assignment.
Users can also select PhysicalAI SimReady material libraries without
downloading the entire collection. The release offers a curated catalog of 265
indexed materials and a full catalog with 1,529 currently usable entries.
More detailed evidence shows which parts were prepared, predicted, assigned a
fallback, successfully bound, or left unbound. Approved assignments are also
preserved when an asset moves from Material processing into Physics processing.
Technical specification text and converted PDF pages are now kept outside the
visual model input. Material Agent selects the image-based material label first;
extracted specification claims can then corroborate that label or flag a
conflict for review, but cannot introduce or replace the visual result.
Generated-material mode is opt-in, limited to single assets, and requires
reference imagery and a configured image-generation backend. Full native
OpenPBR/MaterialX authoring is not included in the default release path.
Physics Agent — Beta
The simulation-tuning workflow is now available as a complete iterative
refinement loop through REST. Applications can run simulation trials, compare
the results with text, image, or video goals, revise the simulation scenario,
and repeat until the result is accepted or the configured limit is reached.
They can then retrieve the tuned USD, any produced recording, and supporting
visual evidence.
The release also improves reference and generated-frame controls, physical
friction constraints, visual judging, durable failure reporting, and ARM64
support for Physics/OvPhysX tuning.
Joint Agent — Research Preview
Joint Agent adds articulation structure to prepared multipart USD assets. It
identifies possible supporting and moving parts and proposes revolute joints
for rotating parts such as doors, lids, wheels, and knobs, or prismatic joints
for sliding parts such as drawers, rails, and telescoping components.
After review, Joint Agent can author the accepted connections, axes, and
movement limits into a self-contained USDZ. For assets requiring aggregate
links or multiple articulation roots, it also authors their exact rigid-link
membership and articulation roots. Package readback verifies that the
published asset contains the approved joint graph. Joint Agent does not create
missing rigid bodies, masses, colliders, drives, or joint states.
For the 17 prepared assets used to qualify this release, all 17 exact packages
passed the selected static Isaac Sim Asset Validator and SimReady Foundation
profiles. The packages contained 78 authored joints, all 119 pinned Foundation
feature checks passed, and no published package differed from its accepted
joint graph.
These results demonstrate static package and SimReady-profile conformance for
the qualified assets. They do not measure inference accuracy on arbitrary
assets or prove dynamically correct motion, contact, travel limits, or
long-term simulation stability.
Texture Agent — Research Preview
Building on its text-to-texture and UV-preparation capabilities, Texture Agent
now supports reference-conditioned, UV-aware editing through an
operator-provided Texture Variation API. Users can target specific materials
or parts, preview a bounded processing plan, and retry individual unsuccessful
texture units while retaining accepted results.
Stricter targeting and a surface-only policy help preserve modeled detail in
CAD, PCB, and industrial assets. Downloaded USDZ packages also retain their
active UV-addressable PBR texture graphs.
The release provides the external-service adapter and API contract. A managed
Step1X runtime, model checkpoints, and setup package are not included; the
default path remains lightweight text-to-texture generation.
Validation Agent — Research Preview
Validation Agent checks render validity, visual appearance, authored physics,
and physical-behavior evidence. Version 0.5 adds direct local OVRTX rendering,
consistent rendering-backend selection, normalized results across supported
model providers, and clear failure when a requested capability is unavailable.
Validation remains available through CLI and Python as a reproducible
release-gating tool.
Agentic Workflow — Research Preview
Fixed pipelines remain the reproducible default. The Agentic Workflow research
preview adds an adaptive mode that can inspect scene structure, intermediate
renders, Validation findings, and previous attempts before deciding what to do
next.
For Material workflows, it can compare rendered results with references,
identify the affected parts, correct only those assignments, rerender the
relevant views, and review the result again. Single-asset workflows support up
to three review and targeted-repair passes by default.
Because one session retains earlier decisions while inspecting each new render,
it can refine fine part colors, logos, and symmetric assignments that a single
fixed pass may miss.
For larger Material scenes, the workflow manages decomposition, per-asset
processing, and collection back onto the original topology. Durable state
allows interrupted or partially failed work to revisit the affected phase
without discarding unrelated completed results.
The preview also supports source conversion, Physics authoring with runtime
Validation, and selected SimReady checks. It does not yet provide agentic
Texture or Joint workflows or general orchestration across all five agents.
Platform improvements
Content Agents 0.5 also strengthens provider selection, credential and artifact
isolation, cancellation correctness, service health and capacity reporting,
configuration diagnostics, container security, and ARM64 support for
Physics/OvPhysX and OVRTX.
Added
- Added opt-in Material Agent generation of asset-specific textured USD
material libraries from reference images and optional text guidance, with
fail-closed backend preflight and single-asset scope. - Added selectable PhysicalAI SimReady material catalogs with lazy category
hydration, including curated and full-catalog options. - Added a Texture Variation service integration for bounded,
reference-conditioned, UV-aware editing of explicitly selected materials and
prims, with targeted regeneration and generated-artifact manifests. - Added the Physics Agent iterative refine REST workflow with progress events,
cancellation, artifact downloads, visual evidence, and configurable
reference and generated-frame sampling. - Added direct local OVRTX rendering for Validation Agent workflows alongside
the existing remote-render path. - Added the Agentic Workflow Research Preview for adaptive visual review, targeted
Material repair, durable large-scene processing, source conversion, Physics
runtime Validation, and selected SimReady checks. - Added an owned structured Joint Rigger core that authors articulation
topology, source-backed limits, diagnostics, and immutable result records
from Joint Agent Stage 2 inputs. - Added self-contained USDZ publication with dependency-closure ch...
Content Agents v0.4.3
Public Content Agents release notes.
This update follows the v0.4.2 public release and covers the delta since v0.4.2.
What's New
- Scene Optimizer Core build resources now come from the public
NVIDIA-Omniverse/usd-optimizeGitHub releasev1.0.3. - Material and Physics service containers can now run local Scene Optimizer from the copied bundle as non-root service users.
- Release notice dependency resolution now follows the pinned release
pyproject.tomlmetadata so notice validation stays aligned with the shipped dependency set. - Scene Optimizer build-resource tests now cover the GitHub release URL shape for both supported Linux bundle platforms.
Content Agents v0.4.2
Public Content Agents release notes.
This update follows the v0.3.10 public release and covers the delta from
v0.3.10 to v0.4.2.
What's New
- Added Validation Agent.
- Added Physics Agent tuning and refinement workflows, including service APIs.
- Expanded Material Agent scene and material pipeline support.
- Updated OVRTX rendering runtime and service integration.
- Added Texture Agent Service shared/S3 storage and multi-instance support.
- Reorganized agent skills into
.agents/skills, with.claude/skillsand
.codex/skillssymlinked to the shared skill tree. - Added deployment collection and Brev deployment assets.
- Refreshed docs, changelog, third-party notices, and tests.