robot_msgs

A cohesive domain model for robot data: the standard message types — ROS's geometry_msgs, sensor_msgs, trajectory_msgs, vision_msgs — as frozen Python dataclasses, with the functions that move between them. Offsetting a pose, interpolating a trajectory, projecting a point, turning a depth image into a cloud: first-class methods, not separate libraries.

• ROS's shape, not ROS. Same packages and field names (the ROS docs are our docs), no ROS install, no codegen. ROS 2 messages and bags are an export (robot_msgs.ros).
• Immutable values over compact numpy storage. Every message is frozen; its arrays are read-only, so slices, rows and columns are free views.
• Runtime dependencies: numpy and the stdlib-only dataclass-values (Value/Space). Plots ([viz]) and ROS export ([ros]) are extras.

A trajectory is one entity with many views

import numpy as np

from robot_msgs.geometry import Pose
from robot_msgs.trajectory import JointTrajectory

t = np.linspace(0.0, 2.0, 201)
traj = JointTrajectory(
    t=t,
    joint_names=("shoulder", "elbow"),
    positions=np.stack([np.sin(t), np.cos(t)], axis=1),
).with_derivatives()

print(traj)                                 # a summary, never 400 numbers
print(traj.positions.shape)                 # batch view, no copy
print(traj[100])                            # one point
print(traj.at(0.123).positions)             # interpolated at any time
for point in traj[:2]:                      # slices are views too
    print(point.t, point.positions)
print(traj.joint("elbow").joint_names)      # one joint, still a JointTrajectory
print(traj.resample(dt=0.5).t)              # onto a fixed grid

grasp = Pose.from_xyz_rpy(0.5, 0.0, 0.2, np.pi, 0.0, 0.0)  # tool pointing down
pregrasp = grasp.shifted(dz=-0.10, local=True)              # back off along the tool
print(pregrasp.position)

JointTrajectory(N=201, t=[0 … 2] s, joint_names=('shoulder', 'elbow'), positions=float64[201×2], velocities=float64[201×2], accelerations=float64[201×2])
(201, 2)
JointTrajectoryPoint(t=1.0, positions=[0.8415, 0.5403], velocities=[ 0.5403, -0.8415], accelerations=[-0.8414, -0.5403])
[0.12268879 0.99243461]
0.0 [0. 1.]
0.01 [0.00999983 0.99995   ]
('elbow',)
[0.  0.5 1.  1.5 2. ]
Point(x=0.5, y=1.2246467991473533e-17, z=0.30000000000000004)

JointTrajectory and PoseTrajectory subclass the domain-free Trajectory (samples at strictly increasing times). Anything else you want to send along rides as channels (traj.with_channel("kp", kp)); a new kind of trajectory subclasses Trajectory and declares its sample() fields — it gets indexing, interpolation, plotting and serialisation for free.

Context ROS has no field for (an origin, units, provenance) rides Header.meta, a frozen str → str mapping that survives every view and file and is dropped on ROS export. A recording whose stamps repeat or step back is built with JointTrajectory.from_samples(..., duplicate_times="drop_later" | "average" | "error"), which returns (trajectory, repair) and writes the same receipt into header.meta; never silently.

Packages

Every message: == and hash by value, a short repr, replace(...), to_dict() / from_dict(), save_npz / load_npz, pickling.

Conventions

SI units · quaternions (x, y, z, w) (MuJoCo/viser wxyz only through xyzw_to_wxyz) · Euler = ROS roll-pitch-yaw (fixed X, Y, Z) · a @ b applies b in a's frame · Header.stamp float seconds, trajectory t relative to it · Header.meta str → str · absent optional arrays are None. The full contract, what we keep from ROS and what we drop, and how to extend: docs/DESIGN.md.

Develop

env -u VIRTUAL_ENV uv run --directory /home/bam/robot_msgs --isolated --extra test python -m pytest -q

Examples: examples/tour.py (above), examples/ur5e_pick.py (waypoints → minimum-jerk joint trajectory → tool pose trajectory through UR5e FK), examples/gallery.py OUTDIR (every plot).