Robot Stereo-hand Coordination for Grasping Cylindric Parts

Radu Horaud, Yves Dufournaud, Long QuanHAL 1996

TL;DR

This paper presents a novel approach for robot grasping of unknown objects using stereo vision and visual servoing. The key innovation is a method to automatically compute grasp set-points from conic features detected in stereo images, enabling precise manipulation without prior object models. This work is crucial for enabling robots to perform complex grasping tasks in unstructured or hazardous environments.

What Problem Does This Paper Address?

Grasping is a fundamental yet challenging task for robots, especially when dealing with objects of unknown shape and size in complex environments. Traditional methods often rely on detailed object models or specific sensor feedback (like tactile or proximity sensors), which can be unreliable or impractical in many scenarios. This paper tackles the problem of achieving robust and flexible grasping by leveraging stereo vision to overcome the limitations of single-camera systems and the lack of object knowledge. The authors aim to enable robots to autonomously determine grasp positions and execute them accurately, even when faced with uncertainties in robot kinematics, object pose, and the environment.

Methods

The core of the paper's contribution lies in its visual servoing approach for grasping. The method involves two main stages:

  • Off-line Grasp Selection: This stage uses stereo vision to extract salient edges from two images of the scene. An operator helps by selecting edges corresponding to a specific part of the object's contour. Conic curves are then fitted to these selected edges in each image. The crucial step here is using these conics to compute the 3D pose of the object. This is achieved by reconstructing the projective cones formed by the conics and the camera's projection center, and then finding their intersection. This intersection, along with the gripper-object relationship (defined by the operator or inferred), allows for the prediction of a desired gripper position in the 3D space. This predicted position is then projected back into the 2D images to serve as the "goal" for the visual servoing algorithm.
  • On-line Grasp Execution (Visual Servoing): In this stage, the robot arm uses visual information to guide the gripper to the predicted goal position in the images. The algorithm minimizes the error between the current gripper position (as seen in the images) and the predicted goal position. This closed-loop control system relies solely on visual feedback, making it robust to errors in robot kinematics or external calibration.

A key technical detail is the pose computation from conics. The paper describes a method to reconstruct the plane containing the conic and then find the intersection of this plane with the projective cones. This process involves solving a polynomial equation to find a specific quadric of rank 2, which represents a pair of planes. The conic is then extracted as the intersection of one of these planes with a projective cone.

Datasets

Not discussed in the paper. The experiments described were performed using a circular calibration pattern and a gas valve.

Key Insights and Ideas

  • Conics as Robust Features: The paper highlights conics as a powerful feature for representing curved and rotationally symmetrical object parts, which are common in many objects requiring manipulation. Their ability to be fitted to ellipses and circles makes them versatile for grasping.
  • Model-Free Grasping: The approach's ability to perform grasping without explicit 3D object models is a significant advancement. By relying on visual features and a defined gripper-object relationship, robots can adapt to a wider range of objects.
  • Stereo Vision for 3D Pose: The use of stereo vision is critical for obtaining the necessary 3D information for pose estimation, which is essential for accurate grasping. This overcomes the limitations of monocular vision for 3D reconstruction.
  • Visual Servoing for Robust Control: The reliance on visual servoing provides a robust control mechanism that is less susceptible to errors in robot kinematics or external calibration, making it suitable for real-world applications.
  • Operator-Assisted Feature Selection: While the goal is autonomous grasping, the initial operator intervention in selecting features (conics) is a practical compromise that simplifies the initial setup and improves reliability.

Key Results

The paper reports successful experimental results in aligning a gripper with a gas valve located approximately 1.5 meters away. The alignment achieved was within 1mm in position and 0.5 degrees in orientation. The method demonstrated robustness, with experiments running over several weeks without re-calibration after an initial setup. The authors also mention that experiments with a circular calibration pattern showed that the circle was reconstructed as an ellipse with an axis ratio of 0.99, indicating good accuracy.

Limitations and Future Work

  • Operator Intervention: The current method requires an operator to select the conic features in the initial images. Future work could focus on automating this feature selection process.
  • Mathematical Rigor: The authors acknowledge the need for a more mathematically rigorous study of the pose algorithm's robustness, particularly concerning its performance even when intrinsic camera parameters are not known with high accuracy.
  • Generalization: While tested on a valve, further experiments would be needed to assess the method's performance on a wider variety of objects with different shapes and textures.
  • Motion Planning and Obstacle Avoidance: The current work focuses on predicting the final gripper position. Future research could incorporate motion planning and obstacle avoidance to enable more complex manipulation sequences.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{horaud1996robot,
  author = {Radu Horaud and Yves Dufournaud and Long Quan},
  title = {{Robot Stereo-hand Coordination for Grasping Cylindric Parts}},
  booktitle = {HAL},
  year = {1996},
}
Figure