This paper, a classic from ICCV 1999, presents an elegant mathematical bridge between the world of uncalibrated computer vision and the mechanics of robotics. It demonstrates that precise robotic control does not require "metric" details—such as camera focal lengths or the exact physical dimensions of a robot's arm—to achieve high precision.
The Challenge of Metric Calibration
In traditional robotics, "metric" information is essential. To move a robot arm to a specific point using cameras (visual servoing), engineers typically rely on a fully calibrated system. This involves knowing internal camera parameters and exact robot link lengths. Calibration is often tedious, and any error leads to inaccuracies. The authors of this paper ask: Can we represent and control complex, articulated motion using only the "weakly calibrated" information available from an uncalibrated stereo rig?
Projective Kinematics
The core of the paper is the development of Projective Kinematics. The authors show that the fundamental algebraic structure of motion is preserved even when viewed through the "distorted" lens of uncalibrated cameras.
- Projective Reconstruction: Using a stereo rig where only the epipolar geometry is known, the system reconstructs the scene in a 3D projective space.
- The PSE(3) Group: Rigid displacements in the real world correspond to a specific subgroup of $4 \times 4$ homographies in projective space, called $PSE(3)$. This group inherits the Lie-group structure of the standard Euclidean displacement group ($SE(3)$).
- Projective Joints: The authors derive mathematical representations for "projective rotations" and "translations." These are parameterized homographies describing how a joint moves in the projective reconstruction.
Non-Metric Visual Servoing
By applying the "Product of Exponentials" formula to these projective joints, the authors derive a Projective Jacobian. This matrix relates joint velocities directly to the velocities of features in the camera images.
This allows for a non-metric control law: a proportional controller that minimizes the error between current and target image features. The system doesn't need to know what a "meter" or a "degree" is; it only needs to know how joint changes affect pixel coordinates. This makes the system inherently robust to changes in camera zoom or unknown robot geometry.
Experimental Validation
The framework was evaluated using a simulated PUMA-like robot arm with six degrees of freedom. The system was tested on its ability to reach a target configuration from an initial one, even with Gaussian noise added to image measurements.
Key Results
- Convergence: The robot successfully reached its target (translating 82cm and rotating 88°) in approximately 20 iterations.
- Precision: With 0.5px of image noise, final joint errors were remarkably small (less than 0.3 units for most joints).
- Stability: The image-space error followed a smooth, exponential decay curve, indicating a stable control law.
Limitations and Future Work
While successful, the system requires an "Acquisition Phase" where the robot performs trial motions for each joint to estimate projective generators. Additionally, the authors noted that while the math is sound, real-world stability requires refining these generators with non-linear numerical methods to handle significant sensor noise. This work remains a foundational bridge between projective geometry and practical robotic control.
Summary updated