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

BEHAVIOR 2026 benchmark server

An implemented, bounded debugging service for BEHAVIOR-1K v3.9.2: one fixed turning_on_radio public instance. A two-control HTTP native smoke check has passed; this is not a full-challenge or model capability result.

• Evaluation protocol: schedule, sensors, actions and scoring.
• Native contract: source evidence and acceptance gates.
• Preparation: licensed assets and runtime prerequisites.

server.py is the only public entrypoint. bench_server/ owns adapter, fixed protocol and private worker communication; environment/ owns native runtime and readiness. Shared HTTP transport remains in benchkit/sim_server. Agent policy and result reports belong in AgentPolicies/<agent>/. Historical tasks.jsonl is a 2025-only inventory, not this service's schedule.

Prepare without starting a simulator

Use a dedicated persistent Python 3.11 / Isaac Sim 5.1 environment and exact source b1979916ec1549b10a4e65e630bc6504a9af1b00. Obtain the separate dataset license before downloading restricted assets. Install the native evaluator dependencies; no model policy server or training weights are needed. Paths are operator inputs.

Run the file-only capture with the native worker interpreter, with this repository and the source's OmniGibson, bddl3, joylo on PYTHONPATH:

python -B -m benchmarks.simBenchmarks.Behavior.environment.deployment \
  --source-root /absolute/behavior-v3.9.2 \
  --data-root /absolute/behavior-assets \
  --receipts /absolute/extraction-receipts.json \
  --output /absolute/new-behavior-deployment.json

The receipts JSON maps behavior-1k-assets, omnigibson-robot-assets, and 2026-challenge-task-instances to extraction receipt paths. Each receipt binds archive_sha256, source_commit, and absolute extraction target. This establishes recorded archive/extraction provenance, not every live asset byte. Selected robot/task files, tracked native package sources, runtime versions, interpreter and module origins are rechecked before initialization. Stale/untracked source bytecode is rejected. Capture never imports OmniGibson or resets a scene.

Launch

Copy server.example.json; replace all paths and the all-zero manifest digest, and select a GPU. A lightweight environment runs the service; worker_python chooses the separate native runtime.

python -B -m benchmarks.simBenchmarks.Behavior.server \
  --parameters /absolute/behavior-server.json --seed 20260922 --episodes 1 \
  --port 8765 --output-dir /absolute/new-run/server

Provide RSI_SIM_SERVER_TOKEN through the environment as described in the shared protocol. Startup/describe do not start SIM. Resetting a claimed case starts one isolated native worker; native shutdown cannot terminate the HTTP process. Use a matching trusted service client.

Existing official assets without old extraction receipts

An explicit deployment_track: "official-live-assets/v1" supports already licensed local assets without claiming their archive/extraction origin. Capture with the native worker interpreter and the same pinned v3.9.2 source:

python -B -m benchmarks.simBenchmarks.Behavior.environment.deployment \
  --deployment-track official-live-assets/v1 \
  --source-root /absolute/behavior-v3.9.2 \
  --data-root /absolute/existing-licensed-behavior-assets \
  --output /absolute/new-behavior-live-deployment.json

Do not supply --receipts for this track. Add the identical deployment_track to the usual server JSON and use the new manifest path and its printed SHA256. Both the service and worker reject a manifest from the other track. The default remains archive-receipts/v1, preserving existing manifest and protocol identities.

The live track verifies the fixed official source, interpreter, module origins, runtime versions, three asset directories and four selected task/robot file hashes before initialization. It records that original archive provenance and the complete asset inventory are not verified. Other live scene/texture assets remain native runtime inputs; a match on four files is not an attestation of all assets. No download, license acceptance or extraction receipt is fabricated.

Its distinct public protocol is behavior-2026-default-rgb-proprio-debug/official-live-assets/v2. DefaultWrapper, 23-value actions, 61-value state, public radio instance 0/native301, timing, initialization, scoring and one-attempt semantics are unchanged. Existing native evidence for the archive track is not evidence for this new deployment track; perform the same two-control infrastructure acceptance before model evaluation.

The actor supplies 23-value direct actions with prefix 1; optional motion guards are policy-owned. There is no expert, planner, oracle or automatic retry. Preserve initialization errors and uncertain execution. Cache is excluded from transfer; logs, controls, private outcome, three videos and cleanup receipts are retained. A two-control diagnostic is not a failed full episode or an SR measurement.

Verified scope

On 2026-09-23, the fixed public instance 0 (native 301) completed initialization, two public-state-derived hold controls and clean shutdown through this HTTP service. Each control advanced four physics ticks (eight total). Three 224×224 RGB cameras and the 61-value robot state were returned; three videos were retained. There were zero model calls. The diagnostic stopped deliberately after two controls, with native_episode_complete: false; its Q=0 is not a model failure or an SR estimate. No other task, full horizon or official RGB-D track was validated by this check.

Tested service revision: bb3c7e99053e669f7ca83a07c94fa9ef337e12ef. Closed diagnostic report SHA-256: 3aeb3ad472d8f3d124c600d295a2ebd2a43d934900eecf0393cece821c7ab36a. Raw no-model evidence remains in operator artifact storage, not participant model results. Every deployment still needs its own readiness check; a passing record is not proof that another host has the same assets, drivers or runtime.