24 October 2026 — shared-action example integration The input-independent task and XR/native adapters are packaged from the exercised shared pick-and-place experiment. The tutorial remains focused on stacking; keyboard/gamepad are optional Lab-native callback integration examples, not new Capture device support. No host credentials, device creation, QA verdict code or host-specific launch paths are included in the package. Prepared-host evidence before packaging/rebase: - Repaired XR feature ceb6941c: stacking 1966 native samples/190 command correlations; pick-and-place 23/23 checks,1268 samples/109 correlations, 0.671 mm placement error. One successful native episode per task. - Native keyboard: 40/40 checks,1771 samples,1.342 mm placement error. - Native gamepad: 41/41 checks,2915 samples,1.774 mm placement error. - Native common adapter: 25 CPU controls and Claude source review accepted. XR four-fix patch: Claude and Codex source reviews accepted. - Keyboard and gamepad initially failed target settling; those attempts remain retained. Bounded correction preserves the original pose tolerances/budgets. - Native keyboard/gamepad used AutoData953cd377 and Lab91a0906f, relative IK; XR used the prepared Capture83e2598e backend and Labcb9628be, absolute IK. XR's installed example wheel and rebuilt client contained the repaired feature. - Primary assistant reviewed sampled video frames, full decode and independent native telemetry/negative controls. No separate model whole-video review. These are single final passes on a prepared x86 host, not reliability, physical sensing or full current-main SDK-build qualification. The packaging/rebase is checked separately; prior recordings are not claimed to execute the new Git head. The packaged/rebased example was then exercised on the same prepared native Capture83e2598e/Labcb9628be backend with the rebuilt client. First motion-only attempt and diagnostic retry failed native export: the callback cached an old scene under a new reset ID on the paused render path. Both failures are retained. The observer now refreshes the scene at each custom-success evaluation boundary. A regression fails against the old package and passes against the fixed wheel. Fresh direct packaged XR pick-and-place after that repair: - One native episode exported; native worker exited normally, task PASS. - 23/23 independent dataset/input/placement checks;1412 native samples, 109 browser/native command correlations;0.310 mm placement error. - Native success began at step1113; final neutral wait required step1194, task checkpoint1196; native export1412 after300 success steps. - Same active episode; every input acknowledgement has a later native step. - Browser-left/Lab-right video decodes cleanly; sampled carry/release frames agree with native telemetry. This is primary review, not whole-video model review. - Fixed destination, export guard and shared task ran from the installed wheel; keyboard/gamepad runtime evidence remains the prepackaging runs above. - Bounded shared-source and reset-snapshot fix reviews accepted by Claude; the earlier four XR guard fixes also had Codex review. Reviews are code-level, not blanket end-to-end certification of every feature or target platform. - 85 installed Python checks,130 client checks,4 browser checks, production/type builds, strict docs and repository pre-commit. Temporary trust and owned processes were cleaned up; unrelated host workloads were preserved. This validates the final example modules on the prepared backend. It does not claim a full SDK rebuild from rebased main or repeated-run reliability. --- Retained end-user installation evidence --- Synthetic XR Teleoperation 0.2.1 — end-user validation, 2026-10-18 Historical native evidence: these runs predate the task-guard corrections for endpoint admission, measured orientation, episode reset and fresh-step action correlation. Those corrections have CPU regression coverage and the prepared-host native recheck recorded below. Do not treat the earlier results as native qualification of subsequent source changes. Executed configuration: - SDK built from our branch refreshed onto upstream main 22026461cceb238fc0f7b3d4aecd3d970d444791. - Isaac Capture 1.7+local, isaacteleop transition distribution 1!1.7+local, CloudXR Runtime 6.3.0, Isaac Sim 6.1.0.0. - Fresh public Isaac Lab checkout cb9628bee016c532feb5fc36c3b012fe707bed1b; fresh uv environment, no imports from the earlier prepared Lab environment. - Ubuntu x86-64, RTX PRO 6000 Blackwell, driver 595.58.03, Python 3.12.14. - Google Chrome 155.0.8059.84: normal namespace/seccomp sandbox verified, exact localhost server certificate trusted; no sandbox/certificate bypass flags. - Desktop IWER, real OOB hub, CONNECT/CDP, native CloudXR and Lab retargeting/IK. - SwiftShader/CPU compositing, H.264, 512x448 per eye; functional scope only. - IsaacContrib-Stack-Cube-Franka-IK-Abs, seed 42, one environment per run. Stacking via the documented public task API: Run 27c78abcefb9443d9e09316c36f7c335: PASS. One native episode, 1453 samples, 200 correlated browser/native commands. Maximum position difference 1.62603325e-07 metres. All recorded actions match native feedback. Final ten native states satisfy independent three-cube geometry. Native environment closed; worker exit 0; no forced shutdown. Full same-run Lab/browser recording: 157.200000 seconds. Custom task acceptance (kept separate from the stacking tutorial): Run 9717f37107884bc3924e9f5839ff6344: PASS. The application uses only FrankaTaskSession and a custom success callback. One native episode, 752 samples, 109 correlated commands. Maximum position difference 1.57192716e-07 metres. Lift 0.159786 m; final XY error 0.000906 m. Last twenty native states verify placement, settling, open fingers and retraction. Native environment closed; worker exit 0; no forced shutdown. Full same-run Lab/browser recording: 125.350000 seconds. Verification and packaging: Both native HDF5 datasets are checked against feedback and command receipts, with native pre-step observations aligned to post-step states. Both videos fully decode. Selected frames show grasp, lift, transfer and release in both views; this is sampled visual inspection, not exhaustive video-model evaluation. At that validation checkpoint, the feature Python modules matched those executed byte for byte. Version, setup documentation and browser-only mock fixtures changed after native runs; production browser source is unchanged. Exact wheel/build identities, final repository checks, independent review and archive checks accompany this package. 25 Python controls, 130 JavaScript tests and four browser checks pass. Retained failures and scope: The first native attempt timed out without advancing/exporting while first-launch Kit/desktop dialogs obscured the session. It was stopped and preserved before resolving the setup prompts and using a new run directory. The long browser-only sweep test exposed a missing decoded mock frame; the real startup watchdog correctly disconnected and released input. Its fixture now supplies that frame. No production timeout or input-release behavior was weakened. Lab's managed uv configuration intentionally overrides several upstream dependency pins; the raw dependency checker reports these conflicts and two excluded optional packages. This is an exercised pinned Lab environment, not a globally conflict-free Python stack. No Error/Traceback lines occur in the two successful native Lab logs. Fresh environment on an existing GPU/desktop host, with download caches reused; not a new operating-system installation or a blinded first-time-user study. The same assistant authored the custom application with prior project context. Earlier driver-crash and MCAP checks remain historical evidence; they were not rerun on 1.7. Repeated reliability, other SDK versions/robots, physical sensing, hand/body/glove devices and real-robot safety remain outside this qualification. Earlier three-episode footage in the tutorial is explicitly labelled historical. This is our branch delivery, not an official NVIDIA SDK release. NVIDIA main must not be modified or merged into as part of this work.