You are on page 1of 5

Mechanics of PR2s Drinking Habits

Project Report Victor Hwang vchwang@andrew.cmu.edu Maheen Rashid maheenr@andrew.cmu.edu

December 4, 2013

Introduction

Pulling a cork out of a wine bottle is a dicult manipulation task. The cork typically has a larger diameter than the mouth of the wine bottle, and deforms to t in to the mouth and neck of the bottle. Therefore the assumptions of rigid body manipulation do not hold. In addition, there are multiple frictional forces between dierent kinds of materials to consider - the friction between a gripper and the cork, a gripper and the glass bottle, and the friction between the cork and the glass bottle. The eect of these frictional forces is dicult to measure and predict. Lastly, the relative positioning of the cork inside the bottle may vary. Corks that are deeply jammed in to the wine bottle may require a dierent approach to uncorking than one used for corks that are not. These factors combine to make the opening of a wine bottle a complex task that is dicult to study without experimental data. On the other hand, a key task for personal robots is to be able to grasp and manipulate objects in ways that are meaningful and helpful to the humans they assist. Opening a wine bottle is one such task. In order to successfully execute the task, the robot needs to know where the bottle is, how to grip it, and what forces and torques to apply to open the bottle. It also needs to know whether or not it has been successful in the task, and of alternate strategies it can use if an approach does not work. Teaching a robot how to open a wine bottle hence involves intelligently using its visual sensors in combination with impedance controls and feedback systems. The aim of our project was two fold: rstly, to gain insights in to what makes opening wine bottles hard, and, secondly, to nd a robust method the PR2 can use to open a wine bottle. We experimented with four dierent approaches the PR2 could use to open a wine bottle: pulling the cork, wiggling and pulling the cork, pulling and twisting the cork up, and pulling and twisting the bottle down. For each approach we diagnosed the possible reasons for its failure, and through this process found both a near optimal method the PR2 can use to uncork a wine bottle and insights in to the mechanics of wine bottle opening.

Implementation

The pipeline for bottle opening consists of a small vision system along with the end eector impedance control package developed by the MIT. The vision aspect consisted of grabbing point cloud data of the bottle, transforming the points into the frame of the hand, and then assuming that the top most 4cm of points corresponded to the cork of the bottle. We compute the centroid of this point cloud and use that as the spatial location of the cork. This framework was implemented using the Point Cloud Library. The trajectory for the end eector controller is specied by providing a time labeled series of stinesses 1

Figure 1: The Coordinate frame for the PR2 gripper. and coordinate frames. Our twist and wiggle motions are specied by setting stinesses at various frames that dier only by orientation (because if the spatial position was changed, the hand would y away from the cork). The controller then attempts to move to this new frame by the allotted time while maintaining the provided stiness during the entire motion. Because of the lack of torque sensors on the PR2, this controller is not closed loop. In merely uses the Jacobian to compute joint torques to apply. As a result, we never know what force or torque is being applied by the arm, especially near joint limits, motor torque limits, and arm conguration singularities. However, we can be sure that the direction of force will be quite accurate.

Experiments

In each of the following experiments, the PR2 locates and holds the bottle with its right gripper and holds the cork with its left gripper. The strength of each grip is not altered during the course of each manoeuvre. Each gripper uses its own frame of reference with the origin dened at the centre of the gripper and the axes dened as illustrated in Fig. 3. The bottle is held in such a way that the vertical axis of the bottle aligns with the z axes of both frames. The forces and torques are dened in terms of these coordinate frames.

3.1

The Pull

In this case force is exerted along the positive z axis for the arm holding the cork. The arm holding the bottle resists. No torque is applied. Not surprisingly, this method performs poorly since the gripper slips o the cork. The grip is not under rst order form closure, and is most vulnerable to slipping in exactly the direction where the force is being applied.

3.2

The Wiggle

In addition to the cork arm exerting force along the z axis, it also alternates between clockwise and anticlockwise twists around the x axis. The bottle arm resists the motion of the cork arm. This method is prone to failure, especially when the cork is set deeply in the wine bottle. The wiggling motion is successful only when both the grippers positioning around the cork is rm to begin with (a lot of contact) and when a small part of the cork is inside the wine bottle.

Figure 2: The grippers are shown in blue, and the cork is shown in red. The deformed cork to the right has less surface contact with the grippers than the straight cork to the right. With each twist, the cork deforms and causes PR2s grippers to lose contact with the corks surface. This is illustrated in Fig 3.2, where the bent cork has less contact with the gripper than the unbent cork. The loss of surface contact causes the eective strength of the grip on the cork to decrease, and increases slippage. At the same time, the corks deformation also decreases the amount of force that is being eectively applied to the base of the cork (or the area of the cork inside the bottle). As the gripper slips up, the deformation increases and causes less and less force and torque to act on the base of the cork.

3.3

The Cork Twist

The cork arm exerts force along the z axis while also twisting along the z axis. The bottle arm resists. The resulting screw motion is quite robust. However it fails when the cork is lodged deep in to the bottle.

3.4

The Bottle Twist

Rather than have the bottle arm resist while the cork arm applies torque and force, we decided to reverse the dynamic and have the cork arm resist while the bottle arm twisted around the z axis while exerting force in the negative z axis direction. This method gave us the most promising results, and works well even when the cork is deeply embedded in the bottle. It is surprising the bottle twist technique works more robustly than the cork twist technique since the magnitude of the applied torque and force is the same in both cases. The static frictional forces acting on dierent media is also the same. However, the radius of the bottle and cork is dierent, so that more force is needed to twist the cork rather than the bottle. The torque and force required to open the wine bottle is close the maximum force the arms can exert. The additional force required to manipulate a smaller radius object (the cork) may not be possible for the controller to provide. As a result of this the eective torque and force exerted by the robot in the cork twist manoeuvre is less than is needed. On the other hand, with the bottle twist method, the torque and friction are closer to the required amount since the object being manipulated has a larger radius. As explained in the Section 2 it is not possible to measure the actual force exerted on the object being manipulated, so it is hard to verify this hypothesis.

Conclusion

We set out to understand what makes the problem of wine bottle opening hard, and how the PR2 should open a wine bottle. The deformation of the cork makes wine bottle opening hard at multiple levels. It can cause the gripper to loose contact with the surface - hence causing less force and torque to be exerted at each point. It can also increase slippage, hence increasing chances of failure.

(a) Pull

The

(b) The Wiggle: PR2 alternates between applying the wrenches shown

(c) The Cork Twist

(d) The Bottle Twist

Figure 3: Dierent bottle opening manoeuvres tried. The blue bars represent the grippers. The dotted line is the Z axis for both the gripper frames. The grey arrows represent resistant force and torque, while the black arrows represent the torque and force applied. At the same time, failure is not always caused by the mechanical properties intrinsic to wine uncorking. Very often the vision system used to identify the location of the cork does not perform very well. This leads the PR2 to grip the rim of the bottle, or to grip an insucient area of the cork, and consequently fail to remove the cork successfully. As discussed in Section 3 The superior performance of the bottle twist over the cork twist may also be indicative of problems related to the controller. The technique that performs the best according to our experiments is the bottle twist technique. However, this does not imply that bottle twisting is in general the best way to open a wine bottle. With grippers (or hands) that are more robust to cork slippage, or with better vision and control systems, it is entirely possible that any of the other techniques we tried could perform just as well, or better. The bottle twist technique is most robust to problems related to the manipulation of the cork, the point cloud vision system, and the controller given our current implementation on the PR2.

Figure 4: The PR2 getting wasted

You might also like