Transports

Transports connect module streams across process boundaries and/or networks.

• Module: a running component (e.g., camera, mapping, nav).
• Stream: a unidirectional flow of messages owned by a module (one broadcaster → many receivers).
• Topic: the name/identifier used by a transport or pubsub backend.
• Message: payload carried on a stream (often dimos.msgs.*, but can be bytes / images / pointclouds / etc.).

Each edge in the graph is a transported stream (potentially different protocols). Each node is a module:

What the transport layer guarantees (and what it doesn’t)

Modules don’t know or care how data moves. They just:

• emit messages (broadcast)
• subscribe to messages (receive)

A transport is responsible for the mechanics of delivery (IPC, sockets, Redis, ROS 2, etc.).

Important: delivery semantics depend on the backend:

• Some are best-effort (e.g., UDP multicast / LCM): loss can happen.
• Some can be reliable (e.g., TCP-backed, Redis, some DDS configs) but may add latency/backpressure.

So: treat the API as uniform, but pick a backend whose semantics match the task.

Choosing a backend

For most users, the important choice is between lcm, zenoh, and shared memory overrides:

• zenoh: the default. Reliable delivery semantics and the same typed message model through LCMEncoderMixin.
• lcm: the legacy path, opt-in. Fast and simple, but UDP multicast is best-effort.
• shared memory (pSHMTransport, etc.): best for large local streams on a single machine.

At the CLI level, you can select the stream transport globally with:

dimos --transport=lcm run unitree-go2
dimos --transport=zenoh run unitree-go2

Zenoh is the default on every platform. Pass --transport=lcm for the legacy multicast path.

Zenoh quickstart

Zenoh ships with dimOS by default (eclipse-zenoh is a base dependency), so there is nothing extra to install.

Zenoh is the default global stream transport on every platform, so nothing below is needed to turn it on — only to turn it off.

What the default talks to

A stock zenoh session is pinned to localhost. It listens on tcp/127.0.0.1:0 and scouts for peers over loopback only, so sibling dimOS processes on this machine find each other and nothing on the LAN can link to them.

Reaching off the machine is opt-in, and each way is independent:

You want	Pass
A robot, dialed directly	--robot-ip 192.168.1.42 (outbound; the listener stays pinned)
Any other peer or a router, dialed directly	ZENOH_CONNECT=tcp/host:7447
Peers discovered across the LAN	ZENOH_SCOUTING=1 (also unpins the listener)
Scouting on one named interface	ZENOH_INTERFACE=wlan0 (also unpins the listener)

Scouting the LAN and refusing links from it are contradictory, so turning discovery outward hands the listener back to zenoh’s all-interfaces default. An explicit listen=[...] on a session always wins over both.

Two ways to override for one run or for your shell:

1. CLI: dimos --transport=zenoh ... or dimos --transport=lcm ... (see CLI for precedence with .env and blueprints).
2. Environment: DIMOS_TRANSPORT=zenoh or DIMOS_TRANSPORT=lcm.

Typical replay (Zenoh is already the default, so no transport flag is required):

dimos --dtop --replay --replay-db=go2_bigoffice run unitree-go2

Architecture notes (Rerun bridge) live under Zenoh in PubSub transports below.

Benchmarks

Quick view on performance of our pubsub backends:

skip

python -m pytest -sv -k "not bytes" dimos/protocol/pubsub/benchmark/tool_benchmark.py

Abstraction layers

We’ll go through these layers top-down.

Using transports with blueprints

See Blueprints for the blueprint API.

From unitree/go2/blueprints/smart/unitree_go2.py.

Example: rebind a few streams from the default LCMTransport to ROSTransport (defined at transport.py) so you can visualize in rviz2.

skip

nav = autoconnect(
    basic,
    voxel_mapper(voxel_size=0.1),
    cost_mapper(),
    replanning_a_star_planner(),
    wavefront_frontier_explorer(),
).global_config(n_workers=6, robot_model="unitree_go2")

ros = nav.transports(
    {
        ("lidar", PointCloud2): ROSTransport("lidar", PointCloud2),
        ("global_map", PointCloud2): ROSTransport("global_map", PointCloud2),
        ("odom", PoseStamped): ROSTransport("odom", PoseStamped),
        ("color_image", Image): ROSTransport("color_image", Image),
    }
)

Using transports with modules

Each stream on a module can use a different transport. Set .transport on the stream before starting modules.

The runnable example below uses a tiny synthetic image publisher instead of CameraModule so it works without a webcam and in CI; the wiring is the same as with a real camera.

ansi=false

import time

import numpy as np
import reactivex as rx

from dimos.core.core import rpc
from dimos.core.coordination.module_coordinator import ModuleCoordinator
from dimos.core.module import Module, ModuleConfig
from dimos.core.stream import In, Out
from dimos.core.transport import LCMTransport
from dimos.msgs.sensor_msgs.Image import Image, ImageFormat

class TickerCameraConfig(ModuleConfig):
    frequency_hz: float = 2.0

class TickerCameraModule(Module):
    """Publish synthetic frames so this example runs without a webcam."""

    config: TickerCameraConfig
    color_image: Out[Image]

    @rpc
    def start(self) -> None:
        super().start()

        def emit(_: int) -> None:
            img = Image.from_numpy(
                np.zeros((480, 640, 3), dtype=np.uint8),
                format=ImageFormat.RGB,
                frame_id="synthetic",
            )
            self.color_image.publish(img)

        period = 1.0 / max(self.config.frequency_hz, 0.1)
        self.register_disposable(rx.interval(period).subscribe(emit))

class ImageListener(Module):
    image: In[Image]

    async def handle_image(self, img: Image) -> None:
        print(f"Received: {img.shape}")

if __name__ == "__main__":
    # Start local cluster and deploy modules to separate processes
    dimos = ModuleCoordinator()
    dimos.start()

    camera = dimos.deploy(TickerCameraModule, frequency_hz=2.0)
    listener = dimos.deploy(ImageListener)

    # Choose a transport for the stream (example: LCM typed channel)
    camera.color_image.transport = LCMTransport("/camera/rgb", Image)

    # Connect listener input to camera output
    listener.image.connect(camera.color_image)

    dimos.start_all_modules()

    time.sleep(2)
    dimos.stop()

13:11:40.135 [inf][ation/worker_manager_python.py] Worker pool started. n_workers=2
13:11:40.776 [inf][/coordination/python_worker.py] Deployed module. module=TickerCameraModule module_id=0 worker_id=0
13:11:40.784 [inf][/coordination/python_worker.py] Deployed module. module=ImageListener module_id=1 worker_id=1
13:11:42.805 [inf][dination/module_coordinator.py] Stopping module... module=ImageListener
13:11:42.809 [inf][dination/module_coordinator.py] Module stopped. module=ImageListener
13:11:42.809 [inf][dination/module_coordinator.py] Stopping module... module=TickerCameraModule
13:11:42.860 [inf][dination/module_coordinator.py] Module stopped. module=TickerCameraModule
13:11:42.861 [inf][ation/worker_manager_python.py] Shutting down all workers...
Received: (480, 640, 3)
Received: (480, 640, 3)
Received: (480, 640, 3)
Received: (480, 640, 3)
13:11:42.862 [inf][/coordination/python_worker.py] Worker stopping module... module=ImageListener module_id=1 worker_id=1
13:11:42.862 [inf][/coordination/python_worker.py] Worker module stopped. module=ImageListener module_id=1 worker_id=1
13:11:42.914 [inf][/coordination/python_worker.py] Worker stopping module... module=TickerCameraModule module_id=0 worker_id=0
13:11:42.914 [inf][/coordination/python_worker.py] Worker module stopped. module=TickerCameraModule module_id=0 worker_id=0
13:11:42.920 [inf][ation/worker_manager_python.py] All workers shut down

See Modules for more on module architecture.

Inspecting traffic (CLI)

dimos spy is the universal transport spy: one live view of every topic moving on every dimOS pubsub transport — names, message rates, bandwidth, sizes, and liveness — whether the system runs on LCM, Zenoh, or both.

dimos spy                     # everything, all transports
dimos spy --transport zenoh   # filter to one transport (repeatable flag)
dimos lcmspy                  # deprecated alias for: dimos spy --transport lcm

dimos topic echo /topic listens on typed channels like /topic#pkg.Msg and decodes automatically:

skip

Listening on /camera/rgb (inferring from typed LCM channels like '/camera/rgb#pkg.Msg')... (Ctrl+C to stop)
Image(shape=(480, 640, 3), format=RGB, dtype=uint8, dev=cpu, ts=2026-01-24 20:28:59)

Implementing a transport

At the stream layer, a transport is implemented by subclassing Transport (see core/stream.py) and implementing:

• broadcast(...)
• subscribe(...)

Your Transport.__init__ args can be anything meaningful for your backend:

• (ip, port)
• a shared-memory segment name
• a filesystem path
• a Redis channel

Encoding is an implementation detail, but we encourage using LCM-compatible message types when possible.

Encoding helpers

Many of our message types provide lcm_encode / lcm_decode for compact, language-agnostic binary encoding (often faster than pickle). For details, see LCM.

PubSub transports

Even though transport can be anything (TCP connection, unix socket) for now all our transport backends implement the PubSub interface.

• publish(topic, message)
• subscribe(topic, callback) -> unsubscribe

from dimos.protocol.pubsub.spec import PubSub
import inspect

print(inspect.getsource(PubSub.publish))
print(inspect.getsource(PubSub.subscribe))

    @abstractmethod
    def publish(self, topic: TopicT, message: MsgT) -> None:
        """Publish a message to a topic."""
        ...

    @abstractmethod
    def subscribe(
        self, topic: TopicT, callback: Callable[[MsgT, TopicT], None]
    ) -> Callable[[], None]:
        """Subscribe to a topic with a callback. returns unsubscribe function"""
        ...

Topic/message types are flexible: bytes, JSON, or our ROS-compatible LCM types. We also have pickle-based transports for arbitrary Python objects.

LCM (UDP multicast)

LCM is UDP multicast. It’s very fast on a robot LAN, but it’s best-effort (packets can drop). For local emission it autoconfigures system in a way in which it’s more robust and faster then other more common protocols like ROS, DDS

from dimos.msgs.geometry_msgs.Vector3 import Vector3
from dimos.protocol.pubsub.impl.lcmpubsub import LCM, Topic

lcm = LCM()
lcm.start()

received = []
topic = Topic("/robot/velocity", Vector3)

lcm.subscribe(topic, lambda msg, t: received.append(msg))
lcm.publish(topic, Vector3(1.0, 0.0, 0.5))

import time
time.sleep(0.1)

print(f"Received velocity: x={received[0].x}, y={received[0].y}, z={received[0].z}")
lcm.stop()

Received velocity: x=1.0, y=0.0, z=0.5

Zenoh

Zenoh provides network pubsub without relying on UDP multicast for the user-facing stream transport. In dimOS it carries the same typed messages by encoding them with LCMEncoderMixin, so existing dimos.msgs.* types still work.

Use Zenoh when:

• you want a transport that behaves better than UDP multicast on macOS
• you are replaying large or high-rate data and want a more reliable network path
• you want to keep the dimOS typed stream model while changing the transport backend

At the stream level, the transport wrappers are ZenohTransport and pZenohTransport. Install, defaults, and CLI versus environment overrides are in the Zenoh quickstart above.

Performance note: zenoh’s session-to-session path (modules in different processes, the common case) benchmarks faster than LCM for small messages and for >=2MiB ones. Delivery within one shared session (co-located modules in one worker) is its slow path for 256KiB-1MiB messages (a few GiB/s); pin shared memory transports for heavy co-located streams. The benchmark has both cases (Zenoh = shared session, ZenohPeers = separate sessions).

The Rerun bridge also follows the global transport: all channels including TF flow over the active backend, and the bridge listens only on it. To additionally bridge external LCM publishers while running Zenoh, pass an explicit pubsubs=[Zenoh(), LCM()].

Per-topic QoS

Zenoh publisher QoS lives on the Zenoh Topic object (see zenohpubsub.py):

skip

from dimos.core.transport import ZenohTransport
from dimos.protocol.pubsub.impl.zenohpubsub import Topic, ZenohQoS

blueprint = blueprint.transports(
    {("image", CameraModule): ZenohTransport(Topic("dimos/image", Image, qos=ZenohQoS(reliability="best_effort", congestion_control="drop")))}
)

When the factory builds transports from the global switch, it applies defaults (default_zenoh_qos in transport_factory.py):

• The agent channels (human_input, agent, agent_idle): reliable, block under congestion (never drop).
• Image/PointCloud2 streams: best-effort, drop under congestion (latest wins).
• Everything else: zenoh defaults (reliable, drop under congestion).

The publisher for a key is declared with the first publish’s QoS. LCM has no per-topic settings, so QoS only applies when transport=zenoh.

Shared memory (IPC)

Shared memory is highest performance, but only works on the same machine.

from dimos.protocol.pubsub.impl.shmpubsub import PickleSharedMemory

shm = PickleSharedMemory(prefer="cpu")
shm.start()

received = []
shm.subscribe("test/topic", lambda msg, topic: received.append(msg))
shm.publish("test/topic", {"data": [1, 2, 3]})

import time
time.sleep(0.1)

print(f"Received: {received}")
shm.stop()

Received: [{'data': [1, 2, 3]}]

DDS Transport

For network communication, DDS uses the Data Distribution Service (DDS) protocol:

skip session=dds_demo ansi=false

from dataclasses import dataclass
from cyclonedds.idl import IdlStruct

from dimos.protocol.pubsub.impl.ddspubsub import DDS, Topic

@dataclass
class SensorReading(IdlStruct):
    value: float

dds = DDS()
dds.start()

received = []
sensor_topic = Topic(name="sensors/temperature", data_type=SensorReading)

dds.subscribe(sensor_topic, lambda msg, t: received.append(msg))
dds.publish(sensor_topic, SensorReading(value=22.5))

import time
time.sleep(0.1)

print(f"Received: {received}")
dds.stop()

Received: [SensorReading(value=22.5)]

A minimal transport: Memory

The simplest toy backend is Memory (single process). Start from there when implementing a new pubsub backend.

from dimos.protocol.pubsub.impl.memory import Memory

bus = Memory()
received = []

unsubscribe = bus.subscribe("sensor/data", lambda msg, topic: received.append(msg))

bus.publish("sensor/data", {"temperature": 22.5})
bus.publish("sensor/data", {"temperature": 23.0})

print(f"Received {len(received)} messages:")
for msg in received:
    print(f"  {msg}")

unsubscribe()

Received 2 messages:
  {'temperature': 22.5}
  {'temperature': 23.0}

See pubsub/impl/memory.py for the complete source.

Encode/decode mixins

Transports often need to serialize messages before sending and deserialize after receiving.

PubSubEncoderMixin at pubsub/encoders.py provides a clean way to add encoding/decoding to any pubsub implementation.

Available mixins

Mixin	Encoding	Use case
PickleEncoderMixin	Python pickle	Any Python object, Python-only
LCMEncoderMixin	LCM binary	Cross-language (C/C++/Python/Go/…)
JpegEncoderMixin	JPEG compressed	Image data, reduces bandwidth

LCMEncoderMixin is especially useful: you can use LCM message definitions with any transport (not just UDP multicast). See LCM for details.

Creating a custom mixin

session=jsonencoder no-result

import json

from dimos.protocol.pubsub.encoders import PubSubEncoderMixin

class JsonEncoderMixin(PubSubEncoderMixin[str, dict, bytes]):
    def encode(self, msg: dict, topic: str) -> bytes:
        return json.dumps(msg).encode("utf-8")

    def decode(self, msg: bytes, topic: str) -> dict:
        return json.loads(msg.decode("utf-8"))

Combine with a pubsub implementation via multiple inheritance:

session=jsonencoder no-result

from dimos.protocol.pubsub.impl.memory import Memory

class MyJsonPubSub(JsonEncoderMixin, Memory):
    pass

Swap serialization by changing the mixin:

session=jsonencoder no-result

from dimos.protocol.pubsub.encoders import PickleEncoderMixin
from dimos.protocol.pubsub.impl.memory import Memory

class MyPicklePubSub(PickleEncoderMixin, Memory):
    pass

Testing and benchmarks

Spec tests

See pubsub/test_spec.py for the grid tests your new backend should pass.

Benchmarks

Add your backend to benchmarks to compare in context:

skip

python -m pytest -sv -k "not bytes" dimos/protocol/pubsub/benchmark/tool_benchmark.py

Available transports

Transport	Use case	Cross-process	Network	Notes
Memory	Testing only, single process	No	No	Minimal reference impl
SharedMemory	Multi-process on same machine	Yes	No	Highest throughput (IPC)
LCM	Robot LAN broadcast (UDP multicast)	Yes	Yes	Best-effort; can drop packets on LAN
Zenoh	Reliable network stream transport	Yes	Yes	Recommended on macOS for heavy replay
Redis	Network pubsub via Redis server	Yes	Yes	Central broker; adds hop
ROS	ROS 2 topic communication	Yes	Yes	Integrates with RViz/ROS tools
DDS	Cyclone DDS without ROS (WIP)	Yes	Yes	WIP