Coordinate Frames and Transformations

Why Does This Exist?

When an RGB-D camera mounted on a tripod spots a coffee mug, it reports the object's position in camera coordinates: for instance, $0.8\text{ meters}$ directly along its optical lens axis ($z_c$), $10\text{ centimeters}$ to the right ($x_c$), and level with the horizon ($y_c = 0$).

If you feed those raw numbers directly to the robot arm's motor controllers, the arm crashes into the table. To the robot arm, "forward" is along its own base axis ($x_B$), "up" is along its vertical column ($z_B$), and the camera is mounted diagonally across the room.

Robots exist in an ecosystem of rigid bodies, each with its own local coordinate system:

• World frame $\{W\}$: The fixed laboratory, warehouse, or room coordinate origin.
• Base frame $\{B\}$: Anchored to the robot's physical mounting plate.
• Camera frame $\{C\}$: Anchored to the optical pinhole focal center.
• Tool / End-effector frame $\{T\}$: Anchored between the gripper fingertips.

Coordinate transformations convert geometric points, trajectories, and sensor readings from the frame that measured them into the frame that must execute motion. Without rigorous transformation algebra, sensors and actuators speak completely different spatial languages.

Basics first: the projective geometry and homogeneous math foundations live in homogeneous-coordinates-and-projective-geometry. Turning camera coordinates into dense 3D point clouds lives in robot-perception-geometry. Mapping tool coordinates into joint angles lives in robot-forward-inverse-kinematics.

Think of It Like This

How It Actually Works

The special Euclidean group $SE(3)$

A rigid body in 3D space can rotate and translate without changing its shape or internal distances. The set of all rigid-body transformations forms the Special Euclidean Group $SE(3)$.

A transformation $\mathbf{T}_{B \leftarrow C} \in SE(3)$ maps a 3D point $\mathbf{p}_C$ expressed in frame $\{C\}$ to its equivalent coordinates $\mathbf{p}_B$ in frame $\{B\}$:

where:

• $\mathbf{R}_{B \leftarrow C} \in SO(3)$ is a $3 \times 3$ orthogonal rotation matrix ($\mathbf{R}^T \mathbf{R} = \mathbf{I}$, $\det(\mathbf{R}) = +1$) whose columns are the unit axes of $\{C\}$ written in $\{B\}$ coordinates.
• $\mathbf{p}_{B \leftarrow C} \in \mathbb{R}^3$ is the translation vector pointing from the origin of $\{B\}$ to the origin of $\{C\}$, expressed in $\{B\}$.

Homogeneous transformation matrices

Writing rotation and translation as separate additions complicates chaining multiple frames (e.g. Camera $\to$ Base $\to$ Wrist $\to$ Tool). Homogeneous coordinates represent a 3D point as a 4D vector $\tilde{\mathbf{p}} = [x, y, z, 1]^T$, turning the affine transformation into a single matrix multiplication:

Mapping a point is simply:

To reverse the transformation and map from $\{B\}$ back to $\{C\}$, we take the matrix inverse. Thanks to the orthogonality of rotation matrices, the inverse of an $SE(3)$ matrix does not require general matrix inversion:

Worked example: camera to robot base transformation

An overhead workspace camera $\{C\}$ observes a tabletop. The camera is rotated $90^\circ$ around the base $x$-axis and translated by $\mathbf{p}_{B \leftarrow C} = [0.5, 0.9, 0.3]^T\text{ m}$.

The rotation matrix $R_x(90^\circ)$ maps camera axes:

• Camera $x_C$ aligns with base $x_B$.
• Camera $y_C$ aligns with base $z_B$.
• Camera $z_C$ (optical depth) aligns with negative base $y_B$.

Construct the $4 \times 4$ homogeneous matrix:

The camera detects a coffee mug at $\mathbf{p}_C = [0.1, 0.0, 0.8]^T\text{ m}$ ($10\text{ cm}$ right, depth $80\text{ cm}$).

Compute the mug position in robot base coordinates $\mathbf{p}_B$:

The mug sits at $\mathbf{p}_B = [0.6, 0.1, 0.3]^T\text{ m}$ ($60\text{ cm}$ forward, $10\text{ cm}$ left, $30\text{ cm}$ above table base).

Verify the round trip using the inverse transform:

The round trip recovers the original camera observation $[0.1, 0.0, 0.8]^T\text{ m}$ exactly.

Code

This Python snippet builds the homogeneous transformation matrix, transforms the camera point to base frame coordinates, and verifies round-trip inversion using NumPy.

1import numpy as np2
3
4def transform_point_se3(5    rotation_matrix: np.ndarray,6    translation_vector: np.ndarray,7    point_c: np.ndarray,8) -> tuple[list[float], list[float]]:9    """Transform point from camera frame C to base frame B, then invert back."""10    t_bc = np.eye(4)11    t_bc[:3, :3] = rotation_matrix12    t_bc[:3, 3] = translation_vector13
14    # Homogeneous point [x, y, z, 1]15    p_c_homo = np.append(point_c, 1.0)16    p_b_homo = t_bc @ p_c_homo17    p_b = p_b_homo[:3]18
19    # Invert transformation: T_cb = inv(T_bc)20    t_cb = np.linalg.inv(t_bc)21    p_c_recovered = (t_cb @ p_b_homo)[:3]22
23    return [round(float(v), 3) for v in p_b], [round(float(v), 3) for v in p_c_recovered]24
25
26r_bc = np.array([27    [1.0, 0.0, 0.0],28    [0.0, 0.0, -1.0],29    [0.0, 1.0, 0.0],30])31p_bc = np.array([0.5, 0.9, 0.3])32pt_cam = np.array([0.1, 0.0, 0.8])33
34base_pt, recovered_pt = transform_point_se3(r_bc, p_bc, pt_cam)35print(base_pt)       # -> [0.6, 0.1, 0.3]36print(recovered_pt)  # -> [0.1, 0.0, 0.8]

Watch Out For

The Quick Version

• Robotic systems maintain multiple coordinate frames: world $\{W\}$, robot base $\{B\}$, camera $\{C\}$, and tool $\{T\}$.
• Rigid body transformations belong to the Special Euclidean Group $SE(3)$, composed of a $3 \times 3$ rotation matrix and a $3 \times 1$ translation vector.
• A $4 \times 4$ homogeneous transformation matrix combines rotation and translation into a single linear matrix multiplication.
• Inverting an $SE(3)$ transformation uses the transpose of the rotation matrix ($\mathbf{R}^T$) and $-\mathbf{R}^T \mathbf{p}$, avoiding numerical matrix inversion.
• Strict notation ($\mathbf{T}_{B \leftarrow C}$) prevents fatal coordinate direction swaps during multi-camera and hand-eye transformations.

What to Read Next

• Degrees of Freedom and Robot Bodies
• Robot Perception: Depth to Usable Geometry
• Forward and Inverse Kinematics
• Homogeneous Coordinates and Projective Geometry