CloudXR + Apple Vision Pro: ~50% of connects permanently wedge xrWaitFrame (pthread_cond_clockwait in omni.kit.xr.core); session never torn down, reconnect impossible until process restart

Summary

In a headless Isaac Lab + IsaacTeleop teleoperation deployment streamed to an Apple Vision Pro via CloudXR, roughly 50% of first client connects permanently wedge the server inside xrWaitFrame. The session is never torn down, so client reconnects cannot renegotiate; only restarting the server process recovers. The other ~50% of connects work normally.

Environment

• Isaac Lab 3.0 generation (isaaclab_physx backend), running headless with --xr
• Kit XR extensions: omni.kit.xr.system.openxr-109.0.0, isaaclab.python.xr.openxr-3.0.0 (+ .headless-3.0.0)
• IsaacTeleop commit 6ee69410efc453b3af0160f5e59fedf260ed941e
• CloudXR runtime launched in-container by the IsaacTeleop launcher; runtime binaries identify as 6.3.0-rc4 (Monado-based OpenXR runtime)
• GPU: RTX PRO 6000 Blackwell Server Edition, driver 580.159.03, Ubuntu 24.04
• Kubernetes pod behind a cloud network load balancer, ports exactly per the Isaac Lab CloudXR teleoperation docs (TCP 48010 + UDP 47998/47999/48000/48005/48008/48012)
• Client: Isaac XR Teleop Sample Client 3.0.1 on Apple Vision Pro, visionOS 26.3.1. Reproduced with CloudXRKit 6.0.3 (stock pin) and 6.2.0.

Symptom

On ~50% of first connects (intermittent; the rest work):

1. Client connects; server logs IsaacTeleop session started; at most a handful of hand-tracking frames arrive, then the pose/tracker path stops permanently (an MCAP recording of the tracker stream stays at 0 bytes).

2. Client display: with CloudXRKit 6.0.3, one frozen (stale) frame; with CloudXRKit 6.2.0 (which fixed the stale-frame-on-connect), a blank view. Same server-side failure either way.

3. The server main thread is permanently blocked inside Kit's XR frame wait. Native stack captured during a live wedge (py-spy --native):

   Thread (MainThread):
       pthread_cond_clockwait (libc.so.6)
       libomni.kit.xr.core.plugin.so (+0x...506, +0x...741)
       libomni.usd.so
       libcarb.eventdispatcher.plugin.so
       libomni.kit.loop-isaac.plugin.so
       libomni.kit.app.plugin.so  (app update)
       ... Python: SimulationContext.render -> app update
   

   GPU utilization drops to ~0%. The wait never returns (observed > 30 minutes).

4. The CloudXR opaque data channel outlives the wedge: client {"type":"ping"} keepalives keep arriving throughout (logged by omni.kit.xr.system.openxr.plugin as Message received from CloudXR has no 'message' or 'payload' field).

5. Client disconnect/reconnect does not recover: the runtime never tears the session down (no teardown log ever appears) and no new session can be negotiated. Only a process/container restart recovers. Application-side teardown is impossible because Kit pumps the XR session state machine from the thread blocked inside the frame wait.

Timeline detail

Headless --xr creates the OpenXR session and trackers ~45–60 s before any client attaches. The wedge occurs during the session/pacing state cycle at client attach.

Mitigations tested (none prevent the wedge)

• NV_ENABLE_POSE_WAIT=0 (present in the shipped avp-cloudxr.env) — wedge still occurs.
• NV_PACER_FIXED_TIME_STEP_MS=33 — verified loaded via env override; wedge still occurred on the next connect with an identical native stack.
• Server runtime already 6.3.0-rc4; client rebuilt on CloudXRKit 6.2.0 — changes the frozen frame to a blank view, does not prevent the wedge.
• Container CPU allocation raised 14 → 48 cores — no effect on wedge frequency.

Possibly related upstream

• Monado MR !2344 (supersedes !1934): frame-sync semaphore left at 0 across session state cycling makes the next xrWaitFrame block forever; fixed in Monado 25.0.0. Is the CloudXR runtime's Monado fork point past this fix?
• Monado issue #492: a wedged per-client IPC thread leaves a zombie session the runtime can never tear down (teardown would run on the wedged thread) — matches the reconnect-impossible behavior exactly.
• isaac-sim/IsaacLab#6214 and isaac-sim/IsaacLab#5575 (hand-pose stream dying while streaming works).

Ruled out

• xrCreateHandTrackerEXT ... XDev does not support hand tracking spam at session start: expected XDev enumeration until the client hand device attaches (per #344). Occurs on healthy sessions too.
• Load balancer / port configuration (matches the docs; the data channel survives the wedge, so transport is alive).

Questions

1. What is the CloudXR OpenXR runtime's Monado fork point relative to Monado 25.0.0 / MR !2344?
2. Which condition variable in omni.kit.xr.core is the main thread waiting on here, and what runtime-side signal is expected to satisfy it? Under what conditions can that wait be unbounded when the client pose/timing stream dies?
3. Can the runtime detect a client whose pose stream died (data channel still live) and force session teardown so the app's frame wait returns with a session-loss event?

Attachments available on request

Full server logs of wedged and healthy connects, py-spy native stack dumps of the Kit process and the CloudXR runtime process captured during a live wedge, and deployment/load-balancer configuration.