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

BEHAVIOR 2026 native preparation

Status: the fixed DefaultWrapper debugging service has passed one two-control HTTP native smoke check; see verified scope. This is not a model score or a guarantee for another deployment. See the launch and file-only deployment capture guide. Do not launch the archived 2025 VLA bridge or publish its task inventory as the 2026 matrix.

Source identity and permissions

Use BEHAVIOR-1K v3.9.2, commit b1979916ec1549b10a4e65e630bc6504a9af1b00, as required by the 2026 evaluation rules. Pin the complete source and installed runtime independently of that mutable web page.

The pinned installer contains separate dataset terms. Code access or an earlier Isaac Sim license acceptance does not establish acceptance of the BEHAVIOR Data Bundle agreement. Obtain operator confirmation before its acceptance flags or restricted data download. Do not redistribute dataset files or keys in the public repository. Use a fresh environment; never run the installer against another benchmark's environment. Native model checkpoints and training demonstrations are not requirements for a fixed-action infrastructure diagnostic.

Native interface to preserve

Read OmniGibson/omnigibson/eval/{eval.py,evaluator.py,r1pro.yaml} and eval/wrappers/ at the pinned commit, not an archived policy launcher. The source-level contract review records the 23-action / 61-state mapping, observation allowlist and remaining runtime checks.

• The official wrapper is RGBDFullResWrapper: head RGB/depth at 720×720, wrists at 480×480, with robot proprioception. DefaultWrapper is a distinct 224×224 RGB plus proprioception debugging configuration without depth; never silently substitute it.
• The bundled robot config uses R1Pro, absolute arm/trunk joint-position controllers, a velocity-controlled mobile base and native gripper controllers. Its assisted grasping setting is an explicit native mechanism, not an invitation to add an RSI task planner. Record it rather than silently changing it. Validate action order, dimension, units and timing against the installed native robot/controller contract before declaring an action schema.
• Native preparation resets, loads the selected instance, settles it for 25 physics ticks while keeping relevant entities still, and resets for rollout. Keep this setup server-private; one rollout does not imply one underlying reset call. No reset/instance search is exposed to the model.
• Native Evaluator.step obtains one supplied action, steps the environment, updates observation/video and native metrics. The service must replace only the policy-input boundary, preserving native termination and scoring. Object states, private task metadata and metric internals remain server-private.

Runtime compatibility and deployment acceptance

CPython crash-handler initialization

The private worker starts with python -X faulthandler -B -m ..., registering the real CPython crash handler before native imports, then enlarges the main thread's POSIX alternate signal stack to at least 64 KiB. Early enable alone is not sufficient: actual AMX use can still prevent crash-traceback delivery on the smaller stack. On an affected Linux x86 runtime, enabling AMX extended CPU state before late faulthandler.enable() can make alternate signal-stack registration fail with OSError: ENOMEM, even with free memory. This is a worker-runtime compatibility fix, not a change to the benchmark. Do not swallow the exception, monkeypatch the handler or disable crash diagnostics. Crash output remains in the private native_worker.log; native source, controls and scoring are unchanged. native_crash_handler.json records the stack sizes and main-thread scope. The Linux x86-64 bootstrap retains its small libc allocation for the process lifetime so a registered signal stack never points at freed Python memory. It does not claim to configure native SDK threads or replace their own crash handlers. See the Linux XSTATE documentation and Python startup options.

Validation sequence

1. Verify a dedicated runtime, licensed asset revision and a fixed public task instance. Record all missing inputs without claiming an attempted rollout.
2. Freeze robot config, native task/instance schedule, observation allowlist and action semantics. Keep the old 2025 inventory separate.
3. The implemented benchkit/sim_server service selects DefaultWrapper RGB plus proprioception, not the full RGBD track. Full RGBD needs a separately reviewed depth representation; never encode metric depth as RGB or silently drop it.
4. Independently review no-reset startup, private-data filtering, native metric bridge, counters, media, source capture and cleanup; run offline regressions.
5. Predeclare one task/instance and two fixed valid controls for a no-model service diagnostic. Preserve any failure, with no automatic retry. Only after this gate may a separate model trial be called executable.

Cold startup may spend many minutes importing the scene and generating collision, shader and rendering caches, especially on shared storage. Monitor CPU, I/O and native logs; do not confuse a live initialization with an acknowledged reset. Choose the bounded worker_timeout before claiming a case. Never extend a frozen run silently or replay a timed-out reset/control.

The recorded smoke check does not grant license acceptance or establish an installation on another host, model capability, or full challenge parity.