ros-pydantic-gen

Point it at a running robot, get typed Python back.

ros-pydantic-gen introspects a live ROS graph through rosbridge and writes a single self-contained module of Pydantic v2 models — one model per message type (nested types included), a topic → model registry, and a snapshot of the parameter server.

# before: what shape is this, exactly?
listener.subscribe(lambda msg: print(msg["pose"]["pose"]["position"]["x"]))

# after: autocomplete, type checking, validation at the boundary
ros_models.subscribe(client, "/odom", lambda m: print(m.pose.pose.position.x))

It talks to ROS over a WebSocket, so it needs no local ROS installation — it runs happily from WSL, macOS, or a laptop pointed at a robot on the network.

---

Contents

• Requirements
• Installation
• Preparing the ROS side
• Quick start
• What gets generated
• Using the generated module
• CLI reference
• Recipes
• How it handles ROS's awkward bits
• Troubleshooting
• Project layout
• Development

---

Requirements

Python	3.9+ (developed and tested on 3.12)
Runtime deps	pydantic>=2.0, roslibpy>=1.5
On the robot	rosbridge_server and rosapi
ROS version	ROS 1 fully supported; ROS 2 works — see ROS 2 notes

Nothing here requires rospy, catkin, or a sourced ROS workspace on your machine.

---

Installation

From the archive

unzip ros-pydantic-gen.zip
cd ros-pydantic-gen

python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate

pip install -e ".[dev]"            # editable install + pytest/ruff

-e (editable) is what you want while the generator is still being adapted to your robot's message set — edits take effect without reinstalling. Drop the [dev] extra if you don't need the test suite:

pip install -e .

Verify

ros-pydantic-gen --version         # -> ros-pydantic-gen 0.2.0
ros-pydantic-gen --help
pytest                             # 81 tests, no ROS required

The console script and python -m ros_pydantic_gen are equivalent; use the latter if you'd rather not install at all:

PYTHONPATH=src python -m ros_pydantic_gen --help

Into an existing project

pip install /path/to/ros-pydantic-gen

Then either call the CLI, or drive it from Python:

from pathlib import Path
from ros_pydantic_gen import ConnectionConfig, RosIntrospector, connect, generate

with connect(ConnectionConfig(host="172.17.0.1", port=9001)) as client:
    graph = RosIntrospector(client).snapshot()

Path("ros_models.py").write_text(generate(graph))

---

Preparing the ROS side

The generator reads the graph through rosapi, a node that ships with rosbridge_suite. It is launched automatically by the standard launch file, so in most cases this is all you need:

ROS 1

sudo apt install ros-noetic-rosbridge-suite
roslaunch rosbridge_server rosbridge_websocket.launch
# listening on ws://0.0.0.0:9090

ROS 2

sudo apt install ros-humble-rosbridge-suite
ros2 launch rosbridge_server rosbridge_websocket_launch.xml

Non-default port (Duckietown, for instance, commonly exposes 9001):

roslaunch rosbridge_server rosbridge_websocket.launch port:=9001

Docker. If the bridge runs in a container, the host reaches it on the docker bridge address — 172.17.0.1 on Linux — provided the port is published (-p 9001:9001). From WSL, the same address usually works when the container runs inside WSL's Docker.

Check the bridge is reachable before generating anything:

python -c "import roslibpy; r=roslibpy.Ros('172.17.0.1', 9001); r.run(); \
print('topics:', len(r.get_topics())); r.terminate()"

---

Quick start

ros-pydantic-gen --host 172.17.0.1 --port 9001 -o ros_models.py

connected to ws://172.17.0.1:9001
  found 33 topics / 23 distinct types
  resolving duckietown_msgs/BoolStamped
  resolving duckietown_msgs/Twist2DStamped
  ...
  found 41 parameters
wrote ros_models.py (57 models, 33 topics)

Progress goes to stderr and the module to the path you gave, so -q silences the chatter for scripting without affecting the output.

---

What gets generated

One module. No package, no imports beyond pydantic. Roughly:

class RosMessage(BaseModel):
    """Base class for every generated ROS message model."""
    model_config = ConfigDict(populate_by_name=True, extra="ignore",
                              validate_assignment=True)
    ROS_TYPE: ClassVar[str] = ""

    def to_ros(self) -> dict[str, Any]: ...


class RosgraphMsgsLog(RosMessage):
    """ROS message ``rosgraph_msgs/Log``."""

    ROS_TYPE: ClassVar[str] = "rosgraph_msgs/Log"
    DEBUG: ClassVar[Any] = 1
    INFO: ClassVar[Any] = 2
    WARN: ClassVar[Any] = 4

    header: StdMsgsHeader = Field(default_factory=StdMsgsHeader)  # std_msgs/Header
    level: int = 0  # byte
    msg: str = ""  # string
    topics: list[str] = Field(default_factory=list)  # string[]


TOPICS: dict[str, type[RosMessage]] = {
    "/odom": NavMsgsOdometry,
    "/rosout": RosgraphMsgsLog,
    ...
}

TOPIC_TYPES: dict[str, str] = {"/odom": "nav_msgs/Odometry", ...}


class RosParams(BaseModel):
    use_sim_time: bool = Field(False, alias="/use_sim_time")
    ...

Every field carries a trailing comment with its original ROS type (# float64[36]), so the generated file doubles as a readable message reference. examples/ros_models.py is a complete one.

---

Using the generated module

Subscribing

import roslibpy
import ros_models

client = roslibpy.Ros("172.17.0.1", 9001)
client.run()

def on_odom(msg: ros_models.NavMsgsOdometry) -> None:
    print(msg.pose.pose.position.x, msg.twist.twist.angular.z)

ros_models.subscribe(client, "/odom", on_odom)

subscribe looks the message type up in the registry, wires up the roslibpy.Topic for you, and validates each incoming payload into a model instance before your callback sees it.

Publishing

twist = ros_models.GeometryMsgsTwist()
twist.linear.x = 0.2
twist.angular.z = -0.4

ros_models.publish(client, "/cmd_vel", twist)

Defaults mirror ROS's own zero-initialisation, so GeometryMsgsTwist() is a valid all-zeros message — no need to spell out every field. That holds for messages with fixed-length arrays too: NavMsgsOdometry() gives you a 36-element zero covariance, not a validation error.

Validating by hand

Useful when you already have a subscriber, or you're parsing a bag dump:

model = ros_models.model_for("/odom")
msg = model.model_validate(raw_dict)          # raises ValidationError on mismatch
payload = msg.to_ros()                        # back to a ROS-shaped dict

to_ros() emits alias names, so reserved-word fields go back out as from, class, and so on — round-tripping is byte-identical.

Type checking

The module is fully annotated. Under mypy or pyright, msg.pose.pose.positon.x (note the typo) becomes a caught error rather than a KeyError at 3am on a moving robot.

---

CLI reference

ros-pydantic-gen [--host H] [--port P] [--secure] [--timeout S]
                 [--include RE] [--exclude RE] [--no-params]
                 [-o PATH] [--no-defaults] [--raw-uint8]
                 [--dump-json PATH] [--from-json PATH] [-q]

Connection

Flag	Default	Effect
--host	localhost	rosbridge host
--port	9090	rosbridge port
--secure	off	use wss:// instead of ws://
--timeout	10.0	per-call rosapi timeout, seconds

Selection

Flag	Effect
--include RE	keep only topics whose name matches the regex
--exclude RE	drop topics matching the regex (applied after --include)
--no-params	skip the parameter server entirely

Output

Flag	Effect
-o, --output	module path (default ros_models.py); parent dirs are created
--no-defaults	emit required fields instead of ROS zero-defaults
--raw-uint8	type uint8[] as list[int] rather than base64
--dump-json	also write the raw graph snapshot to this path
--from-json	regenerate from a snapshot; no connection made
-q, --quiet	suppress progress output

Exit codes: 0 success, 1 connection failure or missing roslibpy.

---

Recipes

Only the topics you care about. A busy robot advertises a lot of noise; regexes keep the generated module reviewable.

ros-pydantic-gen --include '^/(camera|odom|cmd_vel)' -o ros_models.py
ros-pydantic-gen --exclude '(_debug|/rosout)' -o ros_models.py

Capture once, iterate offline. The single most useful workflow when the robot is in a lab and you are not.

ros-pydantic-gen --host 172.17.0.1 --port 9001 --dump-json graph.json -o ros_models.py
# later, on a train, robot switched off:
ros-pydantic-gen --from-json graph.json -o ros_models.py

The snapshot is plain JSON — commit it, diff it, and you have a record of exactly what the robot's interface looked like on a given day.

Strict mode for ingest pipelines. By default fields carry ROS zero-defaults, which makes construction easy but lets a truncated payload validate. When you'd rather catch that:

ros-pydantic-gen --no-defaults -o ros_models_strict.py

Detect interface drift in CI.

ros-pydantic-gen --from-json graph.json -o /tmp/current.py
diff <(grep -v 'Generated :' ros_models.py) <(grep -v 'Generated :' /tmp/current.py)

The Generated : timestamp is the only nondeterministic line; strip it and the output is byte-stable, so a non-empty diff means the interface really changed.

Binary image topics. If you've configured rosbridge with CBOR compression, byte arrays arrive as real lists rather than base64:

ros-pydantic-gen --raw-uint8 -o ros_models.py

---

How it handles ROS's awkward bits

• Nested types — rosapi/MessageDetails returns the full tree, so one call per top-level type pulls in every nested message. geometry_msgs/Quaternion gets a model even though no topic publishes one directly.
• Base64 arrays — rosbridge JSON-encodes uint8[] as base64, so sensor_msgs/Image.data is typed Union[Base64Bytes, list[int]]; you get real bytes in, base64 back out, round-tripping intact.
• Reserved words — from, class, lambda, 2nd_value, schema become from_, class_, lambda_, f_2nd_value, schema_, each carrying Field(alias=...). With populate_by_name=True, either name constructs.
• Fixed-length arrays — float64[36] becomes Annotated[list[float], Field(min_length=36, max_length=36)], so a malformed covariance matrix fails validation instead of propagating. The default is 36 zeros, matching ROS, so NavMsgsOdometry() still constructs.
• Constants — rosgraph_msgs/Log.WARN and friends land as ClassVars, not fields.
• ROS 1 and ROS 2 naming — geometry_msgs/msg/Pose and geometry_msgs/Pose collapse to the same class name.
• Dependency ordering — topological sort; cycles degrade to forward references via from __future__ import annotations.
• Partial failures — a type the bridge can't describe is logged, listed as a comment block in the output, and does not abort the run.

---

Troubleshooting

Could not connect to rosbridge at ws://... The bridge isn't reachable. Confirm the node is up (rosnode list | grep rosbridge), the port is published if containerised, and no firewall sits in between. From WSL, localhost refers to WSL itself — use the container or host IP.

Topics appear in the comment block instead of as models.

# No model could be generated for these topics - the bridge could not
# describe their message type ...
#   /duckie/lane_pose: duckietown_msgs/LanePose

message_details resolves types against the bridge's Python environment, not yours. If duckietown_msgs isn't on the rosbridge node's PYTHONPATH, it cannot describe those types. Fix it on the robot — source the workspace that defines the messages before launching the bridge — and regenerate.

AttributeError: 'str' object has no attribute 'get' This was a real bug, now fixed and regression-tested. roslibpy's ServiceResponse subclasses collections.UserDict, not dict — so an isinstance(response, dict) check silently falls through, and iterating the response yields its keys as bare strings. If you extend the introspection layer, test against collections.abc.Mapping. See tests/test_introspect.py::TestUnwrapTypedefs::test_userdict_response.

Timeouts on a large graph. Each type costs a round trip. On a slow link, raise --timeout 30 and narrow the work with --include.

RosParams values look stale. They are — deliberately. RosParams is a schema whose defaults happen to be the values that were live at generation time. For current values, read the server at runtime with roslibpy.Param(client, name).get().

Validation errors on live data. Usually a genuine mismatch: the robot is running a different build of the message than the bridge described. Extra keys are ignored by design; missing keys are what surface here.

ROS 2. roslibpy's ROS 2 support is still maturing. Message and parameter introspection work; the notable difference is the header type (std_msgs/Header loses its seq field), which the generator reflects faithfully because it reads the definition rather than assuming one.

---

Project layout

src/ros_pydantic_gen/
├── graph.py        # RosField / TypeDef / RosGraph - the domain model
├── introspect.py   # the only module that talks to roslibpy
├── naming.py       # ROS identifiers -> Python identifiers
├── rostypes.py     # ROS primitive -> Python type tables
├── config.py       # ConnectionConfig, GeneratorOptions
├── cli.py          # argument parsing and orchestration
├── codegen/
│   ├── messages.py # one pydantic model per message type
│   ├── params.py   # the RosParams model
│   └── module.py   # section assembly
└── templates/      # boilerplate of the emitted module, as real files

Dependency flow is one-directional: cli → introspect → graph → codegen.

graph.RosGraph is the seam. Code generation never sees a rosapi payload, and introspection never sees a line of generated Python. That separation is what makes --from-json and the offline test suite possible, and it means adding a new output format only touches codegen/.

---

Development

pip install -e ".[dev]"
pytest                    # 81 tests, ~1s, no ROS needed
ruff check .

The suite runs entirely offline against tests/fixtures/fake_ros.py, which mimics rosapi's responses — including the UserDict wrapper, a typedef with fieldarraylen missing, a type that raises, and one that returns nothing.

Generated modules are written to a temp directory, imported for real via importlib, and exercised with realistic payloads. A syntax error or a bad annotation therefore fails a test rather than surfacing on a robot.

tests/test_packaging.py builds an actual wheel and inspects its contents, because PYTHONPATH=src pytest passing tells you nothing about whether the installed distribution works. It needs pip install build hatchling; without them the wheel test skips rather than failing.

If you touch [tool.hatch.build], note that packages = ["src/ros_pydantic_gen"] already ships every file under that directory, templates/*.tmpl included. Adding a force-include for them duplicates each path and hatchling aborts the build.

Extending it. A few natural next steps and where they'd go:

Change	Where
Support services / actions	introspect.py (get_service_*), then codegen/
Emit dataclasses or TypedDicts instead	a new module under codegen/
Change a type mapping	rostypes.py
Adjust the generated boilerplate	templates/*.py.tmpl

When adding a rosapi call, remember the Mapping rule above, and add a fixture case to fake_ros.py so it stays testable without a robot.

---

License

MIT.