MoQROS

Stream ROS 2 topics over Media over QUIC (MoQ), in Rust: images as video, and any other message type as JSON or CDR data.
moqros converts each image to I420, encodes it (VP8, AV1, or H.264, on a GPU when one is available), and
publishes it to a MoQ relay as a
hang broadcast. That's the WebCodecs-friendly format played by the official
@moq/watch web component. Topics of any other type go into a
hang data track instead (see Other message types). It can also go the other way:
subscribe to a broadcast, decode it, and republish it as a ROS topic.
ROS 2 topic ──► moqros-pub ──► moq-relay ──► browser (examples/web: videos + data)
Image / any msg hang QUIC/WebTransport └──► moqros-sub ──► ROS 2 topic
Layout
| Path | What |
|---|---|
crates/moqros |
The library (crates.io, docs.rs). The core (conversion, codec, MoQ publish/subscribe) is ROS-independent; the ros feature adds r2r bridges. |
crates/moqros-cli |
Bridge nodes (crates.io): moqros-pub (ROS → MoQ), moqros-sub (MoQ → ROS), and moqros-bridge (both in one node). Also a colcon (ament_cargo) package. |
examples/image_source |
Sample node that plays a video file (via ffmpeg) or a PNG/JPEG directory as an Image topic. |
examples/sim |
teleop_sim, a tiny simulated robot (rclpy, numpy) for the teleop demo. |
examples/web |
Browser viewer (React, TypeScript, @moq/watch) for several video and data broadcasts. See Web viewer. |
docker/ |
Dockerfile (ROS 2 Jazzy + Rust, moq-relay) and the end-to-end compose.yml. |
End-to-end sample (Docker)
You only need Docker:
On first start, the one-shot samples service generates a test video (samples/sample.mp4) and a PNG
sequence (samples/frames/) with ffmpeg inside the container. Files that already exist are kept. To stream
your own footage, put a video file or a directory of PNG/JPEG frames in samples/ and pass SAMPLE=<name>
(SAMPLE=frames plays the PNG sequence instead of the video). CODEC=av1 streams AV1 instead of VP8.
Open http://localhost:8080 in a Chromium-based browser or Firefox. Both support WebTransport. The page
connects to http://localhost:4443/anon and shows a panel for every broadcast the relay announces: the two
videos sample/image_raw and sample/frames/image_raw, and the latest messages of the data broadcasts
chatter and sample/pose. The hosted viewer connects there too, so it works with the
sample without the web service.

docker/compose.yml describes each service. Besides the viewer, moqros-sub republishes the four
broadcasts on /moqros/image, /moqros/frames, /moqros/chatter, and /moqros/pose.
The relay allows anyone who can reach it to watch and publish. Read Security before you run the sample on a machine others can reach.
The ROS services share one network and IPC namespace, as processes on one robot would. This lets Fast DDS
use shared memory for the 2.7 MB frames. The source publishes with reliable QoS, and moqros-pub runs with
--reliable. With best-effort QoS, large images are split into many UDP fragments, and losing any one of them
drops the whole frame. In testing that cut 30 fps down to 10–20 fps. On a real robot, match the camera
driver's QoS. Both bridges log their message rate every 10 s (message rate).
To check the round trip on the ROS side:
If UDP port forwarding doesn't work in your Docker setup, run the relay on the host instead
(cargo install moq-relay && moq-relay docker/relay.toml).
Teleop demo (Docker)
The other direction too: drive a simulated robot from the browser while watching its camera.
Open http://localhost:8080/?teleop=robot/cmd_vel. The teleop parameter adds a pad to the viewer that
publishes geometry_msgs/msg/Twist commands as a JSON data broadcast. Hold WASD, the arrow keys, or the pad's
buttons to drive, and press Space to stop. Only the page in use (visible and focused) publishes, so with several
tabs or windows open, the one you're using drives. The demo uses the same ports as the sample above, so run one at a time.
| Service | What |
|---|---|
sim |
teleop_sim: a differential-drive robot in a small grid world. It reads /cmd_vel, and publishes a first-person camera (/camera/image_raw), a top-down map (/map/image_raw), and /odom. It stops when no command arrives for 0.5 s, e.g. when the browser tab closes. |
bridge |
One moqros-bridge node for both directions: the two images as the video broadcasts robot/camera and robot/map, /odom as the JSON broadcast robot/odom, and the browser's robot/cmd_vel republished on /cmd_vel. |
relay, web |
The same as in the sample above. |
To watch the commands arrive on the ROS side:
The relay lets anyone who reaches it publish, so anyone who can reach port 4443 can drive this robot. Read Security before you connect a real one.
Installing the bridge nodes
moqros-pub, moqros-sub, and moqros-bridge need a ROS 2 environment, libvpx, pkg-config, and libclang at build time
(see Development).
With cargo, from crates.io:
IDL_PACKAGE_FILTER="std_msgs;sensor_msgs;geometry_msgs;nav_msgs;rosgraph_msgs"
As a ROS 2 package with colcon (crates/moqros-cli has an ament_cargo package.xml):
&&
Bridge behavior
-
QoS: both bridges default to the best-effort
sensor_dataprofile. Pass--reliableto match a reliable publisher, which is recommended for large uncompressed images. -
Latency: if the encoder falls behind,
moqros-pubdrops all but the newest pending image, andmoqros-subalways jumps to the newest group. -
Reconnects: the relay connection redials automatically, and
moqros-subresubscribes when a broadcast comes back. It also resubscribes when the relay retracts a broadcast or routes it to another publisher of the same name, since the relay doesn't end subscriptions then. It stamps republished images with the local receive time. -
Shutdown: on SIGINT or SIGTERM (
docker stop, systemd), both bridges close the relay session before exiting, so the relay ends their broadcasts right away and a restartedmoqros-pubcan announce them again. -
Codec:
moqros-pub --codec vp8|av1|h264. AV1 needs theav1feature and libaom at build time (apt install libaom-dev,brew install aom). H.264 needs theh264feature and Cisco's OpenH264 binary:# downloads from ciscobinary.openh264.orgBrowsers decode all three through WebCodecs; AV1 needs a reasonably recent browser.
-
Hardware encoding: built with the
hwfeature,moqros-pubencodes on a GPU through GStreamer when it can (--hw auto, the default): VA-API on Intel and AMD GPUs, and NVENC on NVIDIA GPUs, for each codec the GPU supports (H.264 on most, AV1 on recent ones, VP8 on few). It tries each known encoder element and keeps the first one that starts and honors keyframe requests, logging which one it uses; with none, it logs a warning and encodes in software, so the codec's own feature is still needed. A hardware encoder that fails mid-stream (e.g. after a GPU reset) is replaced by software, starting with a keyframe, and an encoder that failed at a resolution isn't tried again at that resolution until the process restarts.--hw offalways encodes in software, and--hw requirefails instead of falling back. Decoding (moqros-sub) always runs in software.# Build: GStreamer development files. Run: the plugins with the encoders and parsers, and the GPU driver.gst-inspect-1.0 vaandgst-inspect-1.0 nvcodeclist the encoders GStreamer found; the node needs access to the GPU (/dev/dri/renderD*for VA-API, thevideoorrendergroup). With H.264 on a hardware encoder, OpenH264 is only loaded if the bridge has to fall back to software.
ROS arguments and parameters
Both bridges accept the standard --ros-args: remap the node name (-r __node:=…) and namespace
(-r __ns:=…), and set ROS parameters with -p or --params-file. Each command-line flag can also be
set as a parameter of the same name in snake_case (--keyframe-interval-ms becomes keyframe_interval_ms).
A flag on the command line wins over a parameter, and a parameter wins over environment variables and
defaults. Bool flags also take a value (--reliable=false). Parameters are read once at startup, and
parameters that match no flag are ignored with a warning. In a parameter file, key a namespaced node by
its fully qualified name (/front/moqros_pub); the bare node name only matches in the root namespace.
A repeatable flag (--topic, --broadcast) takes a string array parameter, one element per occurrence
(-p "topic:=[/a, /b]"). If the flag is also on the command line, the command line replaces the whole
array.
The colcon package installs a launch file (moqros.launch.py) and a sample parameter file
(config/moqros.yaml):
The launch file uses $MOQROS_URL when url isn't given, and publishes the broadcast
<namespace>/image_raw (namespace:=/front gives front/image_raw). For a client certificate (see
Authentication), add tls_cert:=… tls_key:=…, or set $MOQROS_TLS_CERT and
$MOQROS_TLS_KEY.
Several topics
One moqros-pub can publish several topics, and one moqros-sub can republish several broadcasts. Repeat
--topic or --broadcast, and add options after the name, separated by commas:
| Bridge | Options | Defaults |
|---|---|---|
moqros-pub --topic TOPIC,… |
format, type, mode, broadcast |
--format and --mode for every topic. The broadcast name is the topic without its leading slash. |
moqros-sub --broadcast NAME,… |
topic, type, frame_id |
--frame-id for every broadcast. The topic is /moqros/<broadcast> (/moqros/image for a lone broadcast). |
moqros-bridge does both in one node and over one relay connection, e.g. for a robot that streams its camera
and takes velocity commands.
It takes --topic like moqros-pub and --broadcast like moqros-sub, with the same options. The flags
that name a single topic or broadcast (--type, and the other side's name) are left out; use the options instead:
Each topic becomes a broadcast of its own. --type and --broadcast of moqros-pub, and --topic and
--type of moqros-sub, name a single topic, so they only work with one --topic or --broadcast. Video
topics start at once. A data topic without a type starts when it appears in the ROS graph. moqros-sub
waits for all its broadcasts at once, and starts each one when it's announced. It skips a broadcast whose
topic it can't create (e.g. a message type this build has no bindings for), and goes on with the rest.
Otherwise, each bridge exits when any of its topics fails.
Other message types
moqros-pub --format json|cdr publishes a topic of any message type. The default, --format video, is for
sensor_msgs/msg/Image.
json: each message is one JSON document, in r2r's mapping of the message fields. Any hang consumer can read it, including a browser; the web viewer shows the latest message. Large arrays (point clouds, images) get big and slow as JSON, so usecdrfor those.cdr: each message is its ROS serialization, passed through untouched. It's compact and lossless, but only ROS-aware consumers can decode it. Images sent ascdrarrive uncompressed and bit-exact.--mode snapshot(default): each message is a MoQ group of its own, and subscribers skip to the newest one when they fall behind. Suits sensor data and state.--mode stream: messages go into one group and are delivered in order, none skipped. Suits logs and commands. The publisher and the relay keep the group's messages, and a subscriber that joins late gets them all from the start of the group. A group holds at most 8192 messages or 32 MB (moq-net's limits); the next message then starts a new group, and late joiners start there. For long-running, high-rate topics usesnapshot.- Type: without
--type,moqros-pubwaits for the topic to appear and asks the ROS graph for its type. Relative topic names resolve against the node's namespace, but remappings aren't applied. The type goes into the catalog, somoqros-subrepublishes with the same type (--typeoverrides it). - Supported types: r2r generates message bindings at build time, and the data bridges only handle types
it generated. With
IDL_PACKAGE_FILTERset, list every package you need (the Docker image includesstd_msgs,sensor_msgs,geometry_msgs,nav_msgs, androsgraph_msgs).
moqros-sub handles both kinds of broadcast: it republishes video as rgb8 images, and data tracks as their
message type. It fails if a broadcast comes back carrying something different from before.
Web viewer
examples/web is a React and TypeScript app built on @moq/watch.
It shows every broadcast the relay announces side by side: a video panel for each image broadcast, and the
latest message of each data broadcast. Its view is kept in the URL, so it can be bookmarked or shared. See
its README for the query parameters, custom data renderers, and development.
The hosted copy at https://yuma-m.github.io/moqros/ connects to http://localhost:4443/anon by default;
set the Relay URL field or the url parameter for another relay. Since it's served over HTTPS, it can only
reach an http:// relay on localhost. Elsewhere, give the relay a real certificate and use https://.
Cloudflare MoQ relay
The bridges and the viewer also work with Cloudflare's MoQ relays, which speak the IETF MoQ Transport drafts (draft-16 here). Create a relay in the Cloudflare dashboard (Media > Realtime > MoQ Relay) or through its API; it comes with a publish/subscribe token and a subscribe-only token. The token goes in the URL path:
For the viewer, put the subscribe-only token's URL in the Relay URL field. Each token covers its whole relay,
so give viewers the subscribe-only one. The bridges hide the token (/REDACTED) when they log the URL, but
Cloudflare's own access logs see it, and so do shell history and process lists; prefer short-lived tokens and
revoke them after testing. The relays are a beta product, so the endpoint and the drafts it speaks may change.
Library usage
The library is published on crates.io as moqros:
[]
= { = "0.4", = ["ros"] } # omit `ros` for the ROS-independent core
use ;
// `node` is an r2r node that you spin.
let publisher = connect?;
let broadcast = publisher.create_image_broadcast?;
let bridge = spawn_ros_to_moq?;
To publish and subscribe over one connection, use moqros::connect_both(&config), which returns a Publisher
and a Subscriber sharing it. See docs.rs for the API, data broadcasts, use without ROS, the codec features
(VP8 by default, AV1 through libaom, H.264 through Cisco's OpenH264 binary, hardware encoders through GStreamer),
and the format on the wire.
Security
The Docker sample is meant for one machine. Before you expose a relay beyond localhost, keep these in mind:
- Anyone who reaches the relay can watch, and publish.
docker/relay.tomlgrants anonymous publish and subscribe underanon/([auth] public = "anon/**"). With port 4443 reachable, anyone who knows the address can watch every broadcast, and can publish broadcasts of their own. Broadcast names aren't secret: subscribers can list what's announced under a path. - Published broadcasts can reach ROS.
moqros-subrepublishes whatever arrives on its broadcasts as ROS messages. On an open relay, someone else can publish under the same name and inject messages into your ROS graph, e.g.geometry_msgs/msg/Twiston a command topic. - Decoded pictures are capped. A few bytes of bitstream can declare a huge picture, so
moqros-subrejects pictures above 8K UHD (7680×4320 pixels) instead of allocating them, and resubscribes. In the library, raise or lower the cap withImageSubscriber::set_max_pixelsorDecoder::set_max_pixels. - The compose file listens on all interfaces. Its
ports:entries bind 4443 (relay) and 8080 (viewer) on every host interface, and on Linux Docker's port publishing bypasses firewalls such as ufw. To keep the sample local, bind to loopback ("127.0.0.1:4443:4443/udp","127.0.0.1:4443:4443/tcp","127.0.0.1:8080:80"). http://relay URLs trust the network. QUIC always encrypts the stream, but with anhttp://URL the client fetches the certificate fingerprint it trusts over plain HTTP, so anyone on the path can swap it and read or alter the stream.--insecureskips verification entirely. Outside development, give the relay a real certificate ([listen] tls.cert/tls.key) and usehttps://URLs. The bridges log a warning at startup for anhttp://URL whose host isn't this machine (localhost,*.localhost, a loopback address), and for--insecure.
Authentication
Replace public in the relay's [auth] section with url, pointing at an auth server: the relay asks it
about every session. moq auth serve from the moq project is the reference
one. It admits a session by a client certificate (mTLS) or a JWT, and grants publish and subscribe patterns.
See the upstream authentication documentation
for its flags. Give each client only what it needs: the robot publishes, viewers subscribe.
Bridges: client certificates. For a device that runs unattended, issue it a certificate from your own CA.
Point the relay's [listen] tls.root at the CA, and pass the certificate and its key to the bridge:
--tls-cert and --tls-key also come from MOQROS_TLS_CERT and MOQROS_TLS_KEY, or the tls_cert and
tls_key ROS parameters; in the library, use ClientConfig::with_client_cert. Nothing secret goes into the
URL, and you revoke a device by its certificate. moq auth serve grants every certificate its CA signed the
same patterns. To scope each device to its own path, run an auth server of your own that decides by the
certificate's name or fingerprint, which the relay passes along with each session.
Viewers: JWTs. Browsers present a JWT in the jwt query parameter of the relay URL. Issue it from your
own backend once the viewer has signed in, signed with the key moq auth serve verifies, and add it to the
viewer's relay URL (the Relay URL field or the url query parameter of examples/web). A token's root is
the path it's valid under, and its publish and subscribe patterns say what it may do there. The viewer keeps the
token out of its address bar, so links copied from it don't carry the token.
The bridges accept a token too (--url "https://relay.example.com/robot?jwt=…"), but prefer certificates:
a URL ends up in shell history and process lists, and the bridges reuse it on every reconnect without asking
for a new token, so once it expires they can't reconnect. They do hide the token (jwt=REDACTED, or a
JWT-shaped path segment as /REDACTED) when they log the URL.
Development
The core library builds and tests natively on any OS. ROS isn't required for it, but the default vp8
feature needs libvpx, pkg-config, and libclang (bindgen generates the libvpx bindings):
# Debian/Ubuntu
# or macOS (libclang comes with the Xcode CLI tools)
# brew install libvpx pkgconf
# H.264 tests (skipped unless OPENH264_LIBRARY is set):
# OPENH264_LIBRARY=$(./scripts/fetch_openh264.sh) cargo test -p moqros --features h264
# AV1 (needs libaom-dev) and the GStreamer backend (needs the GStreamer packages above, plus
# gstreamer1.0-plugins-good, whose software encoders stand in for hardware ones in the tests):
# cargo test -p moqros --features av1,hw
# On a machine with a GPU, also run the checks against its hardware encoders:
# MOQROS_HW_TEST=1 cargo test -p moqros --features av1,hw -- --nocapture hardware_encoders
# Integration test through a real relay:
MOQROS_RELAY_BIN=
The ros feature, moqros-cli, and examples/image_source need a sourced ROS 2 environment and libclang,
because r2r generates its bindings at build time. Use the Docker build stage, or on a ROS machine:
IDL_PACKAGE_FILTER="std_msgs;sensor_msgs;geometry_msgs;nav_msgs;rosgraph_msgs"
License
Licensed under either of Apache License, Version 2.0 or MIT license at your option.
Dependencies have their own licenses. libvpx is BSD-3-Clause, and the env-libvpx-sys bindings are MPL-2.0.
libaom is BSD-2-Clause with the Alliance for Open Media patent license.
GStreamer is LGPL-2.1 and is linked dynamically.
Cisco's OpenH264 binary is covered by Cisco's binary license.