posetree is an object-oriented, thread-safe, programmer-first, Python library for dealing with transforms, poses and frames, designed especially for robotics. It is unique in that Poses know their parent frame and its relationship to other frames which allows for powerful operations with simple code.
pip install posetree
- Write pose math designed for humans to read.
- Think about quaternions as little as possible.
- Never wonder if you are supposed to pre-multiply or post-multiply, or if you forgot an inverse somewhere.
- Standalone, and easy to integrate into any robotics stack (works with ROS but does not depend on it).
from scipy.spatial.transform import Rotation
from posetree import Pose
# Create a pose from perception information
pose_of_tea_bottle = Pose.from_position_and_rotation(
[-0.3, 0.4, 1.2], # Position of the bottle
Rotation.identity(), # Rotation of the bottle
parent_frame="camera", # Parent frame for position and rotation.
pose_tree=pose_tree) # PoseTree object to track frame relationships over time.
# Calculate things based on other frames
height_of_tea = pose_of_tea_bottle.in_frame("world").z
distance_from_gripper = pose_of_tea_bottle.distance_to(gripper_pose)
# Calculate some base motion targets relative to the current robot pose
base_target = base_pose.translate([2, 0, 0]).rotate_about_z(np.pi/4)
base_target2 = base_pose.point_x_at(human_pose).translate([1, 0, 0])
# Get numbers out of a pose to send to a motion API
x, y, _ = base_target.in_frame("map").position
theta = base_target.angle_about_z_to(map_origin)
navigate_to(x, y, theta)posetree takes an opinionated stance on some words that other robotics stacks use somewhat interchangably.
-
Transform: This comes from the verb, to transform. It is how you get from one place to another. It is an operation to change your location from one to another.
- Example: “Take 10 steps forward and then turn 90 degrees to your left”
- Example: "Translate 1m in x, and 2m in y then rotate 30 degrees about the y axis."
-
Pose: This is a noun. It is a physical location and orientation in 3D space.
- Example: “Standing at my front door facing the street.”
- Example: "The location of the left camera lens, z pointing out."
Notice we can take our example pose (standing at my front door facing the street) and transform it into another pose by using our transform instructions (take 10 steps forward and then turn 90 degrees to your left). If we do that we have a new pose that is a bit in front of my house and (in my case) facing our detached garage.
- Frame: This is a pose that is so important we’ve decided to give it a name.
- Example: We could name “Standing at my front door facing the street” “journey_start” and then we could describe that other pose by saying, “from journey_start take 10 steps forward and turn 90 degrees to your left”
- Example: We could name the left camera pose
"camera".
You can sequence Transforms (by multiplying them together, as is traditional). This has a semantic meaning: e.g. "take 10 steps forward and turn left THEN take 1 step backwards and turn around" which is equivalent to "take 10 steps forward and sidestep once right and then turn right". However you cannot sequentially apply positions in 3D space so there is intentionally no multiply operator for Pose. If this seems strange, I promise that as you use this library you will find that nearly every operation is easier and more clearly expressed via in_frame, translate or rotate operations instead of chains of (world_t_robot * robot_t_camera * camera_t_object).inverse() that you might be used to.
When constucting poses it is useful to think of what you expect the pose to be fixed relative to. For example, you might
detect an apple on a table and get a pose from your perception system in the camera frame, but the apple is actually sitting on the table, and we don't expect the apple to move if the camera moves. So you will want to store your pose object with the parent frame equal to odometry/map/world. Then, even if the robot drives around, apple_pose will still refer to the best estimation of the apple's true location (with the usual caveats about localization drift and noise).
Here's how to express that in posetree.
apple_pose = Pose(camera_t_apple, "camera", pose_tree).in_frame("world")Likewise if you have a bin on the back of a mobile robot and you want to define a drop-objects-into-bin pose right above it you can store
that in the robot frame.
drop_pose = Pose.from_position_and_quaternion([-.3, 0.1, 0.25], [0, 0, 0, 1]), "robot", pose_tree)When designing motion APIs with this library, you should be liberal in what frames you accept, and internally convert them to the frame you want to work in.
For example, in a function to move the arm to a pose:
def move_to_pose(self, target_pose: Pose):
# Best Practice: turn Poses into Transforms at the last moment before acting on them.
arm_motion_primitives.move_arm_to_pose_relative_to_base(target_pose.in_frame("robot").transform)This formulation lets you combine the perception outputs and the motion methods for things like this:
def grasp_apple(self, apple_pose: Pose):
# pregrasp is a pose 15 cm above the apple, with z pointing at it.
pregrasp = apple_pose.translate([0, 0, 0.15], frame="world").point_z_at(apple_pose)
self.move_to_pose(pregrasp)
# Move down in the local 'z' of pregrasp until we touch the apple.
self.move_to_pose_until_contact(pregrasp.translate([0, 0, 0.2]))
# If we feel a contact too early, raise some reasonable error:
# We can check the distance using distance_to, even though the parent frames
# of the two poses are different.
if apple_pose.distance_to(get_tool_pose()) > 0.1:
return "Whoops, we probably didn't get the apple."
# Close the gripper and move up a bit from where ever we are, to lift the apple.
self.close_gripper()
self.move_to_pose(get_tool_pose().translate([0, 0, -0.1]))
# Drop it in the bin.
self.move_to_pose(drop_pose)
self.open_gripper()Poses are immutable, and methods like translate return a new pose object. Immutability is nice because it makes them safe to pass into methods and also makes them thread safe, but there is a gotcha to watch out for:
# BAD!!! Do Not Do!
p1.translate([1,0,0])
p1.rotate_about_z(np.rad2deg(90))
p1.with_position_x(5)
# Surprise! p1 has not changed!
# Better
p1 = p1.translate([1,0,0])
p1 = p1.rotate_about_z(np.rad2deg(90))
p1 = p1.with_position_x(5)
# We are replacing p1 with the new pose returned, so this works
# My favorite
p2 = p1.translate([1,0,0]).rotate_about_z(np.rad2deg(90)).with_position_x(5)While a Pose is immutable, the parent frame can (and does!) change over time relative to other frames, meaning that an individual pose can move relative
to other frames. (For example a pose defined in the robot frame will conceptually move as the robot moves relative to a world
frame, even though its position and orientation remain immutably constant.)
To make this very concrete, say the robot starts out at the world origin:
pose_in_robot_frame = pose.from_position_and_rotation([1,2,3], Rotation.identity(), "robot", pose_tree) # fixed to robot
pose_in_world_frame = pose_in_robot_frame.in_frame("world") # fixed to world
pose_in_robot_frame.position # [1,2,3]
pose_in_world_frame.position # [1,2,3]
# Now the robot moves 1 meter forward in the world frame.
robot.drive_forward_in_x(1)
# Poses are immutable so they have not changed. One pose is [1,2,3] from the robot
# frame origin, one is [1,2,3] from the world origin.
pose_in_robot_frame.position # [1,2,3]
pose_in_world_frame.position # [1,2,3]
# But if we express the one in robot frame in the world frame, we see that it
# is now 1 meter farther forward in x (because it was glued to the robot as
# it moved relative to the world.)
pose_in_robot_frame.in_frame("world").position # [2,2,3]If you use ROS2, you can use this excellent adapter library by Brennand Pierce to connect to other modules and visualize poses. https://github.com/brennand/ros_posetree.
To connect a pose_tree instance you need to subclass PoseTree. Lets say your stack uses a (fictional) object called MyTransformManager that subscribes to pose messages and implements get_transform that returns some flavor of transform struct. You would write something like:
class MyPoseTree(CustomFramePoseTree):
"""My implementation of PoseTree to integrate with MyTransformManager"""
def __init__(self, transform_manager: MyTransfrormManager):
self._tfm = transform_manager
def _get_transform(self, parent_frame: str, child_frame: str, timestamp: Optional[float] = None) -> Transform:
transform_data = self._tfm.get_transform(parent_frame, child_frame, timestamp)
return Transform.from_position_and_quaternion(
[transform_data.tx, transform_data.ty, transform_data.tz],
[transform_data.qx, transform_data.qy, transform_data.qz, transform_data.qw]
)Then you can use it like this:
my_tf_manager = MyTransformManager()
my_tf_manager.subscribe(Channels.ALL_ROBOT_CHANNELS) # idk I'm making this API up.
pose_tree = MyPoseTree(my_tf_manager)
p1 = Pose.from_position_and_rotation([1,2,3], Rotation.identity(), "robot", pose_tree)
p2 = p1.in_frame("camera", timestamp = robot.get_current_timestamp() - 1.0)For more detailed information about the API and how to use PoseTree, check out the documentation.
We welcome contributions! This is my first open source project so I don't know what I'm doing, but I'm excited to help you figure out how to incorporate it into your projects.
posetree is licensed under the MIT license.