Geometric Robot Calibration Using a Calibration Plate

AI summary

Overview

Research area: Robotics — geometric calibration of robot and machine kinematics using a low-cost measurement artefact.

Technical level: Advanced. The paper works with extended forward kinematics, Taylor-series linearization of a position error vector, and constrained nonlinear optimization, and assumes familiarity with kinematic frames, rotation matrices, and least-squares parameter identification.

Scope: The paper introduces a calibration plate with precisely known inter-point distances as a substitute for absolute measurement systems such as laser trackers, and validates it experimentally on a three-axis laser cutting machine.

What This Paper Is About

Robotic and gantry systems deviate from their ideal kinematic models because of mounting and manufacturing inaccuracies, so the kinematics must be expanded with error parameters that are then identified from many measurements. Conventional approaches measure the absolute position of the end effector with a laser tracker or motion capture system, which is expensive and mechanically delicate. This paper instead measures the relative positions between points on a calibration plate whose distances are precisely known in advance, and uses the mismatch between measured and known relative distances to identify the error parameters.

Key Contributions

1. A calibration concept based on relative rather than absolute measurement. The plate's known inter-point distances in its body-fixed frame M replace absolute position measurements of the end effector, so a laser beam is only used to position the machine at the center of each four-quadrant diode target.
2. A calibration plate design and its integration into the kinematic loop. The forward kinematics is extended by the non-measurable laser beam length L to close the loop down to the point of impact on the sensor surface; a selection matrix S_xy is introduced to eliminate the laser lengths when working in the end effector frame, producing 3(n-1) equations per pose.
3. Two identification procedures for the error parameters. A linearized least-squares formulation separating intrinsic (int, constant across poses) and extrinsic (ext, per-pose plate orientation and laser lengths) parameters, and a constrained nonlinear optimization with an explicit objective function and lower/upper bounds on parameters.
4. Experimental validation against a laser tracker. The method is tested on a three-axis laser cutting machine and compared directly with a LEICA AT930 Absolute laser tracker calibration.

Main Findings

• Both identification methods agree. The least-squares approach and the constrained optimization "led to approximately the same results." The constrained optimization was preferred and used for the reported experiments because the constraints prevent estimated values from exceeding the assumed realistic range.
• Large uncalibrated errors. The uncalibrated forward kinematics r_0E_I(q) deviates from the laser tracker reference r_0E,ref_I by absolute errors up to some millimeters; the tabulated maximum Δr_max,xy is 3.27 mm and the mean Δr_mean,xy is 1.87 mm.
• Substantial improvement from the proposed plate method. Using the calibration plate, the maximum x-y error fell to 0.81 mm and the mean error to 0.26 mm.
• Laser tracker calibration is more accurate but close. The laser tracker calibration achieved a maximum x-y error of 0.54 mm and a mean of 0.11 mm, so the plate method's results are described as correlated with one of a laser tracker calibration.
• Mean error reduction of about 86%. Comparing the uncalibrated case with the kinematics corrected by the identified error parameters r_0E_I(q, p_e), the mean error was reduced by about 86%.
• Only x and y were evaluated. Because the error parameters related to the z direction are difficult to identify given the machine and calibration plate geometry, the error was evaluated only in the x and y directions, which the authors note is what matters for laser cutting machines anyway.

Methodology in Plain English

The researchers place a plate carrying four-quadrant diode targets somewhere inside the working area of a gantry machine. A laser pointer on the end effector is steered into the center of each diode, and the axis encoder values are recorded at each measuring point. Because the diode target positions are known precisely inside the plate but the plate's own placement and orientation on the work area are unknown, the plate orientation γ_M and the laser beam lengths L become additional unknown quantities alongside the machine's geometric error parameters p_e.

The model is closed by extending the forward kinematics from the end effector to the actual laser impact point on the sensor surface. Since the laser length cannot be measured, one option is to transform the position difference between two measuring points into the end effector frame and apply a selection matrix that removes the z component where the lengths live — this yields 3(n-1) equations per pose but loses one equation and can fail for certain parameter choices. The alternative used for the final identification builds the position error vector in the plate's body-fixed frame M, keeping the laser lengths as parameters to be identified in exchange for one more equation per measuring point.

The error parameters are then estimated in one of two ways. The least-squares approach linearizes the error vector with a Taylor series, assembles matrices Q and Theta over all m calibration poses (splitting intrinsic and extrinsic contributions), and iterates p_id^(n+1) = p_id^(n) - [Theta^T Theta]^{-1} Theta^T Q until the parameter change falls below a chosen border. Initial error parameters are set to zero (or small random values on the order of the expected errors where needed), laser lengths are initialized from the uncalibrated kinematics, and plate orientations are estimated roughly from how the plate sits on the work area. The second approach instead minimizes a nonlinear objective f = Δr^T Δr / 2 subject to lower and upper bounds p_min ≤ p_id ≤ p_max on error parameters, plate orientation angles, and laser lengths; this is implemented with CasADi using the IPOPT solver and the same initial guess.

For validation, a three-axis laser cutting machine with an integrated laser beam was used, and a raster laser tracker measurement over the whole workspace was taken with a LEICA AT930 Absolute system as an independent reference.

Why This Matters

Impact on research. The work shows that geometric calibration does not necessarily require absolute end effector position measurement; relative distances between points of a precisely characterized artefact can carry enough information to identify model error parameters, bringing calibration closer to camera-calibration practices such as separating intrinsic and extrinsic parameters.

Real-world applications:

• Initial calibration of gantry machines, including milling machines, 3D printers, and laser or water cutters, which the introduction lists as beneficiaries of a well-known system description.
• Laser cutting machines specifically, where the authors note x-y accuracy is the practically relevant quantity.
• Fast recalibration after a crash, which the conclusion describes as "a straightforward process" with this method.
• Field or shop-floor calibration where transporting and setting up a laser tracker is impractical, thanks to the plate's small size.

Industry relevance. The plate is described as mechanically robust, cheaper than conventional reference devices such as laser trackers or motion capture systems, and easy to transport because of its small size. The achievable accuracy nevertheless depends directly on how precisely the reference distances d_ik_M on the plate were determined in advance.

Future Directions

• Extension to industrial robots. The conclusion names the extension of the method for industrial robots, including their orientation, as further research; currently the modeling and identification are demonstrated for a gantry machine.
• Making z-direction error parameters identifiable. The paper reports that z-related parameters were difficult to identify given the machine and plate geometry, and notes that certain error parameters require the calibration tool to have different sensor heights to become identifiable.
• Designing a plate setup for those parameters. This points to tailored plate geometries — for example, varied sensor heights — as an open design question rather than a fixed artefact.
• Choice of constraints for the optimization. The authors caution that constraints must be chosen in a plausible range because solutions outside the restrictions are not taken into consideration, leaving the question of how best to select realistic bounds open.

Target Audience

Researchers and engineers working on robot and machine tool calibration, kinematic modeling, and parameter identification; practitioners in laser cutting, milling, 3D printing, and gantry machine operation who need an affordable in-house calibration route; and readers interested in camera-calibration-style intrinsic/extrinsic parameter separation applied to mechanisms rather than cameras.

Authors’ abstract

In this paper a new method for geometric robot calibration is introduced, which uses a calibration plate with precisely known distances between its measuring points. The relative measurement between two points on the calibration plate is used to determine predefined error parameters of the system. In comparison to conventional measurement methods, like laser tracker or motion capture systems, the calibration plate provides a more mechanically robust and cheaper alternative, which is furthermore easier to transport due to its small size. The calibration method, the plate design, the mathematical description of the error system as well as the identification of the parameters are described in detail. For identifying the error parameters, the least squares method and a constrained optimization problem are used. The functionality of this method was demonstrated in experiments that led to promising results, correlated with one of a laser tracker calibration. The modeling and identification of the error parameters is done for a gantry machine, but is not restricted to that type of robot.