TL;DR
This paper presents a novel multi-sensor combination measuring system (MCMS) that fuses data from visual and angle sensors to significantly improve the pose accuracy of robot manipulators. By employing Kalman filtering (KF) and multi-sensor optimal information fusion algorithm (MOIFA), the system achieves pose accuracy improvements of 38%-78% without requiring complex kinematic calibration or motion constraints.
What Problem Does This Paper Address?
Industrial robots often require high pose accuracy for tasks like precision assembly, welding, and manipulation. However, traditional methods for improving this accuracy, such as kinematic calibration, are often performed offline, are sensitive to environmental changes, and can be computationally intensive. Existing online methods also have limitations, including complex calibration steps or requirements for large, open spaces. This paper addresses the need for a more autonomous, efficient, and accurate method to improve robot manipulator pose accuracy in real-time.
Methods
The core of this work is the development of a Multi-Sensor Combination Measuring System (MCMS). This system integrates:
- A visual sensor (photogrammetry system): Used for real-time measurement of the manipulator's position. The system utilizes four motion-sensitive CCD cameras.
- An angle sensor (digital inclinometer): Rigidly attached to the manipulator to measure its orientation (pitch and yaw).
- A series robot: The robot itself provides its own pose information.
To fuse the redundant data from these sensors, two data fusion methods are employed:
- Kalman Filter (KF): A recursive algorithm that estimates the state of a dynamic system from a series of noisy measurements. It's well-suited for real-time applications due to its predictive capabilities.
- Multi-Sensor Optimal Information Fusion Algorithm (MOIFA): A geometric fusion method that aims for minimum variance unbiased estimation.
The paper also details a robot calibration method using the photogrammetry system to establish a transformation relationship between the photogrammetry system's coordinate frame and the robot's base coordinate frame. This involves modeling kinematic parameter errors and using a least-squares method to correct them.
Datasets
The paper does not explicitly mention pre-existing datasets for training. Instead, it relies on:
- Simulated data: Random points were generated within a 100x100x100 mm space, with errors added based on typical instrument accuracies to simulate real-world measurements.
- Experimental data: Collected from a KUKA KP 5 arc robot in a controlled lab environment. The photogrammetry system and a FARO Xi laser tracker were used for measurements.
Key Insights and Ideas
- Synergy of Visual and Inertial Sensing: Combining a visual system (photogrammetry) for position with an angle sensor (inclinometer) for orientation provides complementary data that significantly enhances pose accuracy.
- Data Fusion is Key: The effectiveness of the MCMS hinges on sophisticated data fusion techniques like KF and MOIFA, which can leverage redundant sensor information to reduce overall error.
- Optimal FOV for Visual Sensors: The repeatability of the photogrammetry system is highly dependent on its Field of View (FOV). An optimal range of 1x0.8x1 m to 2x0.8x1 m was identified for best accuracy.
- Calibration Simplification: The proposed method avoids complex kinematic parameter equation solutions and motion constraints often found in traditional vision-based methods, making it more autonomous and efficient.
- KF vs. MOIFA Trade-offs: KF offers real-time prediction, crucial for dynamic applications, but its accuracy can degrade with larger measurement ranges. MOIFA provides stable fusion accuracy but lacks predictive capabilities, potentially leading to hysteresis in feedback control.
Key Results
- Pose Accuracy Improvement: The MCMS with data fusion achieved dramatic improvements in robot manipulator pose accuracy, ranging from 38% to 78%.
- Calibration Effectiveness: After calibration using the photogrammetry system, the robot's position error was reduced to less than 1 mm.
- Simulation Results:
- KF reduced error by 78.2%.
- MOIFA reduced error by 46.1%.
- KF showed smaller fused errors than MOIFA in simulation.
- Experimental Results:
- KF reduced error by 67.3%.
- MOIFA reduced error by 38.2%.
- The experimental results generally mirrored the simulation trends, though the error reduction percentages were slightly lower.
- Optimal FOV: The best repeatability accuracy for the photogrammetry system was found in the range of 1x0.8x1 m to 2x0.8x1 m, with average errors of dx=0.124 mm, dy=0.997 mm, dz=0.272 mm, and da=1.045 mm.
Limitations and Future Work
- KF and Measurement Range: KF, being a linear filter, can experience enlarged fused errors with increasing measurement ranges. This was noted as a limitation and a point for future verification.
- MOIFA Hysteresis: MOIFA's lack of predictive function can lead to hysteresis in feedback control systems.
- Future Work: The authors suggest future work will focus on real-time data transferring, online control, and compensation for the robot manipulator's pose. They also propose exploring the trade-offs between KF and MOIFA based on specific application requirements (dynamic vs. static, small vs. large measurement ranges).
Summary updated