Category: Omniverse / Isaac Sim
Title: Isaac Sim livestream on TCP-only clouds (AutoDL / RunPod / etc.): keep WebRTC local, view via noVNC
TL;DR
Isaac Sim’s official WebRTC livestream cannot cross cloud proxies that only forward TCP/HTTP: media is SRTP over UDP (47998). Setting publicIp does not fix a missing UDP path, and current 5.x/6.x releases do not expose a documented TURN “TCP-only” knob.
Workaround that preserves NVENC livestream: run Isaac streaming and the official WebRTC Streaming Client on the GPU host (UDP stays local), then view the client window through TurboVNC + noVNC on the single TCP port your provider already gives you.
Scripts + write-up: https://github.com/0xHoseaHAO/isaac-sim-tcp-only-cloud
(Tested: Isaac Sim 6.1, AutoDL-style Linux container, RTX-class GPU.)
The problem
On TCP-only GPU clouds:
- Signaling on TCP
49100may connect - Media on UDP
47998never reaches a laptop behind the HTTP proxy - Result: grey/black viewport
This matches the official docs (both TCP and UDP required) and earlier forum reports about TURN/ICE not being configurable for TCP-only setups in 5.x.
What we did instead
Browser --TCP--> noVNC --> TurboVNC desktop
|
v
Official WebRTC Streaming Client
|
UDP localhost
|
v
Isaac Sim streaming + NVENC
isaac-sim.streaming.shwithprimaryStream/publicIpset to the container bridge / eth0 IP (reachable by the local client — not the public HTTPS proxy URL).- Official WebRTC Streaming Client on
DISPLAY=:1(TurboVNC), auto-connected via CDP. - Expose only noVNC (default
:6006) through the cloud custom HTTP service.
How this differs from “GUI into noVNC”
There is already an excellent RunPod write-up that puts the full Isaac GUI into Xvfb/noVNC. That is a great path when it works.
On Isaac 6.1 we often could not get a stable Vulkan present surface on plain Xvfb. Running official streaming locally keeps hardware NVENC and the supported livestream path, while still using one TCP port for the browser.
Extra gotcha: NVENC on multi-GPU containers
If the host has many GPUs but the container only sees /dev/nvidiaN with N != 0, NVENC may fail with “unsupported device” on newer drivers. The repo includes a small LD_PRELOAD interposer (nvenc_fix.c) that filters GPU enumeration to devices that actually exist under /dev.
Ask / hope for upstream
A documented ICE/TURN (or TCP-relay) configuration for livestream would help TCP-only clouds a lot. Until then, local-WebRTC + noVNC is a practical pattern for AutoDL / RunPod / similar providers.
Happy to share logs or refine the scripts if this is useful to others.