common/source/starBenchserver/benchmarks/simBenchmarks/RoboDojo/environment/README.md

RoboDojo environment

This directory owns the native Direct v2 worker, robot-only kinematics and hash-bound runtime preparation. It contains no model policy, prompt, provider client, training recipe or model result.

Task definitions and fixed panels live in ../bench_server/protocol/task_catalog.json. The service adapter uses native_worker_client.py to communicate with this directory's worker. That private connection is separate from the public SIM service used by model runners. Deployment paths and GPU selection are supplied through ../server.example.json.

Required runtime

For a fresh deployment from official public sources without the historical private parent manifest, use official-source deployment v1 and set deployment_track: "official" in the existing server configuration. The historical deployment described below remains available unchanged.

Use legally obtained RoboDojo source/assets and the exact revisions recorded in reference_eef_snapshot.json. The accepted deployment baseline uses Python 3.11, Isaac Sim 5.1, IsaacLab 0.54.3, Torch 2.7.0+cu128, NumPy 1.26.0, SciPy 1.15.3, Gymnasium 1.2.1 and PyAV 18.1.0. These are validation pins, not an automatic installer.

The operator must provide a compatible NVIDIA driver/GPU and accept applicable runtime and asset licences. The tools do not copy an EULA marker, credentials or licensed assets. Do not fabricate a manifest, weaken source checks or silently replace unresolved materials.

Prepare an owned deployment

The create-only preparation tool copies verified source and assets into a new runtime root and emits an immutable manifest/digest:

python -B -m benchmarks.simBenchmarks.RoboDojo.environment.reference_eef_runtime prepare \
  --native-repository /absolute/upstream/git/repository \
  --base-assets /absolute/qualified-parent/Assets \
  --additional-assets /absolute/authorized-reference/Assets \
  --reference-assets /absolute/authorized-reference/Assets \
  --asset-parent-manifest /absolute/qualified-parent-manifest.json \
  --runtime-root /absolute/new-owned-runtime \
  --python /absolute/native-environment/bin/python \
  --isaaclab-root /absolute/pinned/IsaacLab \
  --curobo-root /absolute/pinned/curobo \
  --recording-support /absolute/recording-support \
  --archive /absolute/new-owned-runtime.tar.gz

Asset locations are operator-supplied paths. --base-assets supplies the exact parent files inventory; --additional-assets supplies the pinned additional asset leaves. The optional --reference-assets independently rechecks the parent manifest's reference_files at a current location, including its absent counterparts and declared URDF line-ending equivalence. It may point to the same tree as --additional-assets, but the two options verify different inventories. When omitted, preparation still verifies every base and additional asset byte; it does not access the historical counterpart directory. The preparation receipt records whether the optional counterpart comparison ran and its resolved path.

The parent manifest's reference_root is historical provenance, not a hardcoded deployment requirement. Keep the original parent manifest bytes and pinned digest when relocating assets. The service reads the prepared runtime's Assets tree via its configured runtime manifest; it does not need the original reference directory. No asset-root field is passed to the model.

KNOWN_UNRESOLVED records exact inherited material/texture references inside the pinned assets. Those URI strings identify accepted missing references; they are not deployment paths or configurable asset roots.

Retain the returned manifest, digest, preparation receipt and archive hash. External interpreter/dependency paths are part of the deployment binding. A relocated installation needs a new verified manifest; never edit a manifest already associated with an episode.

Readiness gate

Before model evaluation verify:

1. source, asset, layout and dependency hashes;
2. Isaac licence acceptance and actual GPU availability;
3. native reset and all three RGB streams;
4. measured robot state, fixed DLS conversion and one acknowledged control;
5. native scorer accounting, video decode and acknowledged worker cleanup.

A reset-only or hold diagnostic is infrastructure evidence, not task success or a model score. Store diagnostics outside public capability results. Never add simulator ticks while waiting for a model: waiting must not change physics.

See the full Direct v2 binding and the parent evaluation protocol.