common/source/starBenchserver/benchmarks/simBenchmarks/Robotwin/README.md

RoboTwin

This benchmark-owned server exposes native Aloha joint-target control, three RGB cameras, language, 12 measured arm joint angles and two explicitly labeled gripper command values. The model generates all task actions. The server does not expose expert rollouts, object poses, IK or task-planning methods.

The adapter is a development fast track, not the native full score: choose lift_pot or click_bell, one predeclared unscreened seed, one attempt. Native imports, reset, image correctness and compiled controller validation are runtime acceptance gates; a listening server or CPU test is not a native pass. See EVALUATION_PROTOCOL.md for exact differences.

The current state14/fast-v3 and OIDN-compatible state14-oidn233/fast-v4 protocols add this robot-state reference for absolute joint control. The former RGB-only v1/v2 results keep their frozen contracts; they do not gain retrospective state. Arm positions come from native measured-joint getters, not the misleadingly named joint_action drive targets. Gripper values remain commands, not measured opening or grasp evidence. The new observation track needs its own native acceptance; an old reset-only or zero-action model run is not that proof.

Directory responsibilities

server.py is the public factory/launch entrypoint; bench_server/adapter.py owns the supported two-task interface and native scoring. environment/ owns deployment, native/runtime pins and renderer readiness; checks/ contains offline regressions. The small task definition remains in the adapter: no empty catalog or protocol package is required. Operator paths and track selection belong in server.example.json or an operator-owned copy. No root adapter shim remains; this relocation does not change native actions, budgets or scoring.

Native environment and assets

The inspected local RoboTwin archive has no own Git revision. Its selected runtime closure is pinned in environment/native_snapshot.json: 2,088 files / 1,011,813,984 bytes, including Aloha, both object variants for the two tasks, native task/config code and its CuRobo package. native_source_revision is null; the descriptor uses a content-inventory SHA-256, not an invented upstream or parent-repository commit.

This initial archive-specific adapter is not yet a one-command clean upstream installation. An arbitrary current RoboTwin clone or the public asset release alone will not satisfy the source pin. Supply the exact matching archive, or review and version a new source inventory/protocol explicitly. Do not regenerate the accepted inventory merely to bypass a mismatch.

The intended native environment is Linux/Python 3.11.15, SAPIEN 3.0.0b1, MPLib 0.2.1 and TOPPRA 0.6.3, with the exact versions and SAPIEN/MPLib/TOPPRA package bytes in environment/runtime_snapshot.json. The adapter checks them without importing the simulator. This is not a full transitive environment, GPU driver or CUDA JIT attestation. Dependencies, renderer/GPU availability and matching asset bytes are separate requirements. In particular, bundled OIDN 2.0.1 has a limited CUDA architecture range. Run the renderer readiness check on the allocated device: valid files and nonempty RGB frames do not establish a working denoiser. For CUDA devices outside that version's architecture range, an explicit oidn_2_3_3_cuda_v1 renderer track is available. Follow the owned overlay instructions and use server-oidn233.example.json. It keeps the original environment and native task bytes untouched. The new renderer track has a different protocol identity and does not claim image parity with 2.0.1; the default example still selects the original track.

To materialize a new deployment from a matching reference archive:

python -B -m benchmarks.simBenchmarks.Robotwin.environment.deployment prepare \
  --reference-root /absolute/path/to/matching-reference-archive \
  --output-dir /absolute/path/to/new-owned-robotwin-deployment

This copies the selected closure without changing the reference. Only the two Aloha curobo_left.yml / curobo_right.yml files are derived from their pinned native _tmp.yml templates by native ASSETS_PATH expansion. The receipt records original/template/deployed hashes and the deployment_manifest_sha256. Every other byte and the exact file set must match. Extra source overrides, symlinks and bytecode caches are rejected. Do not run a native wholesale installer against an existing environment: its package patches, path rewrites and unpinned downloads are not this deployment.

The closure includes assets/objects/objaverse/list.json and assets/objects/same.json: native utility imports read both even when clutter is disabled. They are unchanged native metadata, not generated substitutes. These two clean tasks do not use randomized background/clutter geometry. That does not establish release completeness or enable the randomized/full track. Leave at least the selected closure's size plus headroom for a new copy, native JIT caches and model-owned media. Full RoboTwin assets are substantially larger.

Start one SIM service

From the RSI repository root, copy server.example.json to an evaluator-private configuration. Replace both placeholder deployment fields; the all-zero digest is intentionally not runnable. Select the task and native seed before the attempt. The shared harness seed is recorded separately and does not search for a native seed.

First verify files/runtime only:

python -B -m benchmarks.simBenchmarks.Robotwin.environment.runtime \
  --root /absolute/path/to/new-owned-robotwin-deployment \
  --manifest-sha256 YOUR_DEPLOYMENT_MANIFEST_SHA256

Set the same evaluator-only RSI_SIM_SERVER_TOKEN (32–256 URL-safe characters) in the server and evaluator environments. The adapter creates a fresh per-episode cache/curobo-jit and scopes TORCH_EXTENSIONS_DIR there during native calls; it stays outside the pinned deployment and is excluded from artifact transfer. Using the native environment, start:

python -B -m benchmarks.simBenchmarks.Robotwin.server \
  --parameters /absolute/private/robotwin-server.json \
  --seed 20260907 --episodes 1 --port 8771 \
  --output-dir AgentPolicies/anonymous_code_agent/results/robotwin/raw/lift_pot-pot-lifting-service

-B is required. The server binds loopback and starts with zero native resets. Calls are serialized; waiting for a model does not advance physics. Follow the shared service setup for evaluator-only URL/token setup and SSH forwarding when the evaluator is on another node. Never give service credentials or evaluator-side native records to the actor.

Run a participant through the evaluator

Use a fresh model-owned result directory and a previously prepared opaque policy mailbox; model setup belongs under AgentPolicies/<profile>/, not here. For the external code-agent mailbox, the evaluator command is:

python -B -m benchkit run \
  --benchmark robotwin \
  --sim-server http://127.0.0.1:8771 \
  --model-profile anonymous_code_agent \
  --policy-args '{"directory":"/absolute/opaque-mailbox"}' \
  --seed 20260907 --episodes 1 \
  --budget-mode horizon --max-policy-calls 200 --replan-interval 5 --max-steps 1000 \
  --output-dir AgentPolicies/anonymous_code_agent/results/robotwin/raw/lift_pot-pot-lifting-evaluation

The native task still stops at 400 absolute joint-target controls, or native success earlier. The runner's separate 200 calls / prefix 5 / 1,000-control ceiling does not extend that horizon. Native controls, action decisions and variable 250-Hz physics ticks are different counters. Policy persistence rules are not benchmark requirements. No automatic reset, step replay or retry occurs.

Videos are named head_camera.mp4, left_camera.mp4, right_camera.mp4 under the model's result namespace, never here. PNG/hash records preserve delivered RGB bytes; 10-fps MP4s present control-boundary samples, not real-time physics or a 10-Hz controller. frames.jsonl records physical tick counts. After result transfer, stop this one-episode server process: native close_env alone does not attest GPU-memory reclamation. An errored attempt remains an attempt.

Regression and remaining scope

python -B -m unittest discover \
  -s benchmarks/simBenchmarks/Robotwin/checks -p 'test_*.py' -v

Tests exercise factory, boundaries, single-attempt lifecycle, counters and content-pinned deployment without native reset. Optional array tests need NumPy. Native reset plus fixed infrastructure controls must be separately reviewed and recorded as non-capability diagnostics before participant use. Randomized settings, the native 50-task/accepted-seed schedule and serial/sharded native parity remain unimplemented. The separate official track below provides upstream bootstrapping for the two declared clean tasks.

Retired StarVLA model integration code remains in Git history, not as a server dependency or supported evaluation entrypoint.

Official-source deployment without the historical archive

Use server-official.example.json to select the additive official source/runtime track. It shares the same server entrypoint and public qpos14 interface. Its source, assets and runtime identities are separate from the historical archive; see the additive track in the protocol.

1. Download official RoboTwin commit ea8b21121ebb3cd201ff5b3fe361944ac94eda3f and the two pinned official asset archives. Retain the download receipt with repo_id, revision, archives: [{name, sha256}], and sorted extracted files: [{path, size_bytes, sha256}]; asset file paths start with assets/.

2. In the selected Python 3.10.12/SAPIEN 3.0.3 runtime, capture installed package bytes without importing the simulator:

   python -B -m benchmarks.simBenchmarks.Robotwin.environment.official capture-runtime \
     --output /absolute/new-runtime-snapshot.json > /absolute/runtime-capture.json
   

3. Copy the verified source/assets into a fresh owned deployment. assets-root is the directory containing assets/:

   python -B -m benchmarks.simBenchmarks.Robotwin.environment.official prepare \
     --source-root /absolute/official-RoboTwin-checkout \
     --assets-root /absolute/downloaded-asset-root \
     --asset-receipt /absolute/asset-receipt.json \
     --runtime-snapshot /absolute/new-runtime-snapshot.json \
     --output /absolute/new-official-deployment
   

4. Put its returned manifest digest and root in the official example parameters, then launch the existing rsi-bench serve --benchmark robotwin route with python -B. The existing native diagnostic gate still applies before policy evaluation; preparation and hashes alone are not a native evaluation.