Workpiece Orientation Correction with a Robot Arm Using Visual Information

TL;DR

This paper presents an algorithm for a robot arm to automatically correct the orientation of a workpiece it's holding, using visual feedback. This is crucial for enabling robots to handle un-oriented parts, significantly expanding their industrial applicability and reducing the need for precise pre-orientation.

What Problem Does This Paper Address?

The industrial use of robots is often limited by the need for precisely oriented workpieces. If robots could automatically adjust the orientation of a workpiece they've acquired, even if it's not perfectly positioned, their versatility and applicability would dramatically increase. This paper tackles the challenge of developing a method for a robot arm to refine its estimate of a workpiece's pose (position and orientation) using visual information, allowing it to move the workpiece along a pre-defined safe trajectory without collisions.

Methods

The core of the algorithm is based on linear perturbation theory. The authors model the relationship between changes in visual features of the workpiece and the corresponding changes in the robot's pose. This relationship is captured by a Jacobian matrix, which is determined during an "instruction phase."

  1. Instruction Phase: The robot arm is moved through a series of known perturbations (by adjusting arm joints) at specific "nominal poses" along a desired trajectory. For each pose, visual features (like centroid, area, eigenvectors) are extracted, and the corresponding changes in these features due to the perturbations are measured. This data is used to compute the Jacobian matrix for that nominal pose.
  2. Execution Phase: Once a workpiece is acquired, its initial estimated pose (hTw) is refined. The algorithm compares the current visual features of the workpiece to those recorded during instruction. If the differences are within a "linear region" (meaning the linear perturbation model is still valid), the algorithm calculates a correction transformation matrix (Ap) using the pseudo-inverse of the Jacobian. This correction is applied to update the hand-to-workpiece transformation. The process iterates until the deviations are within a tolerance zone, ensuring a refined pose estimate.
  3. Collision Avoidance: To ensure the workpiece moves safely, the refined pose is compared to the "legal way" the workpiece was held during instruction. If the deviation is small, it's assumed no collisions will occur.

Datasets

Not discussed in the paper. The method relies on visual information extracted from TV cameras during an instruction phase and then during execution.

Key Insights and Ideas

  • Linear Perturbation for Pose Refinement: Using linear perturbation theory to relate visual feature changes to pose changes is a clever way to make pose estimation computationally tractable and iterative.
  • Two-Phase Approach (Instruction/Execution): Separating the learning of the Jacobian (instruction) from the actual correction (execution) allows for a robust and efficient refinement process during operation.
  • Jacobian as a Key Component: The Jacobian matrix acts as the crucial link between the visual world and the robot's physical actions, enabling precise corrections.
  • Handling Uncertainty: The algorithm is designed to handle approximate initial estimates of workpiece pose, making it practical for real-world scenarios where perfect initial acquisition is difficult.

Key Results

The paper describes the algorithm and its theoretical underpinnings. It states that the "selection of visual features and procedures to increase the rate of convergence... are issues still to be resolved." The authors indicate that "The status of experimental results will be reported at the conference," implying that experimental validation was ongoing or planned. Therefore, specific quantitative results are not presented in this document.

Limitations and Future Work

  • Convergence Rate: The authors acknowledge that increasing the rate of convergence for the pose estimation is an area for future work.
  • Visual Feature Selection: Choosing the most effective visual features for robust pose estimation is identified as an open problem.
  • Kinematic Configurations: The arm joint solution procedure is noted as being specific to different kinematic configurations, suggesting that generalization to various robot arm types might require further development.
  • Experimental Validation: The paper focuses on the algorithmic development, with experimental results to be reported elsewhere. The practical performance and robustness in diverse real-world scenarios are not detailed here.

Summary updated

BibTeX
@inproceedings{1977workpiece,
  title = {{Workpiece Orientation Correction with a Robot Arm Using Visual Information}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1977},
}
Figure