marble

Marble

For generated-site Go1 inspection, use one workflow: workflows/testing/marble-go1-inspection.yaml with exactly one native --preset warehouse|shipyard|home|factory|utility selection. Omission defaults to warehouse. Each submission runs one hosted generation using WLT_API_KEY, one RTX PRO 6000 collection and one continuous 25 FPS standalone HTML report. It never fans out across situations. The guide is docs/workbench/guides/marble-robot-presets.md; for a custom world, omit --preset and override world_prompt and workflow_preset on the same YAML. Use the native write_quadruped_gallery to combine completed reports into an offline situation selector retaining every original video and trajectory row. All five parameterizations have prior native GPU evidence from the retired combined workflow; distinguish that evidence from new-entrypoint submissions. These are Go1 presets, not ANYmal or Spot. Do not advertise thermal inspection, defect detection, household manipulation or new RL training.

For detailed quadruped collection, use workflows/testing/marble-warehouse-quadruped.yaml and docs/workbench/guides/marble-warehouse-quadruped.md. Reuse an existing generated warehouse: no new World API request is needed. The runtime verifies pinned official Unitree Go1 meshes and a Google MuJoCo Playground ONNX policy. PyBullet torque dynamics and ONNX inference execute on CPU; gsplat RGB, Warp depth, and Cycles CUDA articulated rendering use one RTX PRO 6000. OpenImageDenoise runs on CPU; CPU path tracing is rejected. This is pretrained-policy collection, not new RL training or a physical robot connection. Preserve upstream licenses, actual link poses, per-foot contact forces, raw depth, and per-frame renderer evidence in the standalone HTML and collection bundle. Verify the completed run and browser playback before claiming live acceptance.

For GPU learning beyond rendering, use workflows/testing/marble-navigation-rl.yaml and docs/workbench/guides/marble-navigation-rl.md. The workbench.marble.navigation_prepare adapter preserves a generated world's collision triangles, measures supported cases with disjoint goal locations, and seals inputs for the existing native ANYmal/Isaac/RSL-RL trainer. Training and evaluation use standard Workbench/SkyPilot stages and immutable S3 handoffs. Require an operator-qualified immutable Isaac image; do not default to the quarantined public release. This adapter's CPU preparation is verified against the real warehouse, but native GPU PPO and held-out evaluation remain unverified. The earlier sensor collection cannot establish navigation learning acceptance. WLT_API_KEY is needed for upstream generation, not for reusing its world bundle.

For embodied warehouse collection, use workflows/testing/marble-warehouse-rover.yaml and docs/workbench/guides/marble-warehouse-rover.md. It requires real World API generation and aligned mesh/splat transforms. workbench.marble.rover_collect drives a four-wheel rover through PyBullet CPU rigid-body dynamics, then renders RGB with gsplat CUDA and raycasts depth with Warp CUDA on one Nebius GPU. It saves actions, joints, contacts, poses, raw depth, and a standalone HTML replay. The observer uses a CPU-rendered rover at recorded poses and CUDA mesh depth occlusion. Do not describe this as physical Spot collection or RL training. Run native runtime submission from a Linux operator and preserve its isolated SkyPilot identity for resume and cleanup.

For a straightforward user walkthrough, use docs/workbench/guides/marble-to-nebius.md: World Labs generates the scene, the CPU acquisition stage downloads files to S3, and a Nebius GPU worker loads those files for rendering or scanning. The report replays completed outputs.

For manufacturing, use workflows/testing/marble-manufacturing-pallet-detection.yaml. It requires WLT_API_KEY and a real pallet dataset manifest, fails before generation on missing auth or invalid/overlapping data, and never substitutes the sample world. Native submit automatically resolves and forwards this token from the environment or private NPA credentials. See docs/workbench/marble-manufacturing.md for the fully headless command. The experiment renders factory backgrounds on CUDA, composites training-only RGBA pallet cutouts, and invokes the existing Faster R-CNN trainer/evaluator on CUDA. Both arms receive matched optimizer-update counts and use the same held-out real evaluation partition. Report negative AP changes honestly. This is a 2D background-augmentation experiment, not a physics simulation or production accuracy claim. New manufacturing execution is unverified until a funded API credential and the operator's real dataset are supplied.