embodiedgen

Files (1)

• SKILL.md

SKILL.md (first lines)

# EmbodiedGen V2 image-to-rigid-object

Use this skill only for the scoped, Tier-1 EmbodiedGen V2 workflow. The stable
user surface is [`workflows/testing/byof-embodiedgen.yaml`](../../../workflows/testing/byof-embodiedgen.yaml)
through generic `npa workbench workflow` commands—not a persistent service,
separate SDK submitter, or an `embodiedgen` command family. Start with the
[operator guide](../../../docs/workbench/embodiedgen.md), then use the
[workflow catalog](../../../workflows/testing/README.md#bring-your-own-framework)
and [OSS capability record](../../../docs/workbench/oss-solution-catalog.md#embodiedgen-v2).

## Capability boundary

The qualified private capability is
`img3d-cli_trellis_image_to_urdf_pybullet`. The real upstream TRELLIS backend
runs through EmbodiedGen `img3d-cli`, emits a URDF and collision meshes,
converts that URDF to MJCF, and validates the exact generated URDF in PyBullet
for gravity contact, settling, and a fully decoded MP4. TRELLIS is selected;
SAM3D and the Tencent cloud backend are not substitutes.

Generated mass, height, and friction are VLM estimates—not calibrated ground
truth. Do not claim articulated-object generation, physics validation in
MuJoCo/Genesis, downstream policy improvement, or model quality. MJCF is a
supported handoff only; PyBullet is the simulator actually exercised.

## Before submission

1. Resolve the project, private registry, storage, and Kubernetes context from
   supported NPA configuration. A host name is not a project alias.
2. Run `npa workbench health preflight --checks nebius,token_factory,s3 --json`
   with the selected project, then `npa workbench health access --capability
   token_factory --json`. TRELLIS is public: do not invent an acceptance flag.
3. Obtain an operator-private, immutable
   `REGISTRY/npa-embodiedgen@sha256:<64-hex>` image. The tag before resolution
   must be `dev-<full-40-character-NPA-sha>` when an operator builds a new
   image. Consumers use the already-qualified digest from supported configured
   private delivery; never publish or use a public registry image.
4. Pass `NEBIUS_TOKEN_FACTORY_KEY` only with `--secret-env`; it is needed by
   upstream property estimation and must never be written to YAML, reports, or
   shell logs.
5. Use a worker-readable `https://` or `s3://` image URI, never a local path.
   The workflow base64-transports it through the rendered shell as literal data;
   reachability and storage permission remain separate prerequisites.

## Operate through Workbench

Run `validate-spec`, `plan-spec --check-render`, `submit --runtime --plan-only`, target
`preflight-images`, then the real `submit --runtime` with one matching `base_image`
config value and
`--image-override workbench.byof.repo=<same-digest>`. Reuse the same project,
run ID, workflow S3 prefix, input URI, bucket, and image digest for status,
logs, artifacts, cancellation, and any `--resume-run` reconciliation. Do not
use raw SkyPilot commands or create a parallel controller.

`plan-spec --check-render` is a local static render and accepts no
`--config-path`. Use the same task-scoped `--config-path` for image preflight,
submit, and resume. Its private SkyPilot configuration must select the intended
context, the standard `skypilot-service-account`, and an existing
private-registry pull Secret; see the [operator guide](../../../docs/workbench/embodiedgen.md#private-registry-target-configuration).

The workflow declares and requires these actual S3 outputs in addition to the
summary: `generated_asset.tar.gz` (URDF plus mesh layout), `mjcf_asset.tar.gz`,
`pybullet_view.png`, `pybullet_settle.mp4`, and
`embodiedgen_image_to_rigid_object.json`. Read the primary JSON and its hashes
before unpacking a bundle or handing it downstream. Use generic
`npa workbench workflow artifacts` or `npa studio search` for discovery; the
generic `npa.sdk.workbench.workflow` SDK is read-side only.

For interruption, inspect durable status/logs/artifacts first. The workflow has
no TRELLIS mid-generation checkpoint; a verified retry can rerun the stage but
must not be described as model-level resume. Cancel only the exact run and use
`npa cleanup --json` to audit local residue. Never clean up a shared controller,
cluster, cache, registry, or another owner's resources.

## Delivery and qualification

EmbodiedGen source is Apache-2.0; TRELLIS source and
`microsoft/TRELLIS-image-large@25e0d31ffbebe4b5a97464dd851910efc3002d96` are
MIT and runtime fetched. The CUDA bootstrap's delivery boundary and component
inventory are in `npa/docker/workbench/embodiedgen/REDISTRIBUTION.md` and
`THIRD_PARTY_NOTICES.md`. Runtime fetch does not grant any source, model, input,
output, or service right.

One exact private digest completed the standard RTX PRO 6000 workflow and all
six declared outputs were read back with 59 hash-verified artifacts. The
operator-private evidence has a historical source bridge covering protected
image/runtime/workflow/profile inputs; later guide and readiness handoff edits
are reviewed separately. It is not a public release or a claim that an
arbitrary target is ready. Every new selected target still needs health/access
checks and exact-image pull preflight, while a new source/image still needs the
private build and qualification procedure in the operator guide.

From nebius/nebius-physical-ai, Apache-2.0.

More from nebius/nebius-physical-ai