Desired Functionality
- - Robot detects the pose of a human’s sword using head camera
- - Robot holds a sword firmly using one of his gripper
- - Robot moves his arm to block the attacker’s sword and defends himself
- - Robot plans a path avoiding the attacker’s sword and fights back
- - Robot performs the action of pulling out a sword
Design Criteria
- - Robot needs to detect the artag using its head camera consistently
- - The size of ARTag should be proper for the robot to detect in his reachable range
- - User can use the cuff button to control the open and close operation of the gripper
- - The position and orientation of ARTag we get is as accurate as needed
- - Robot can execute a path meeting the accuracy requirements
- - Constraints can be added to improve the robustness of the system
Design Illustration
The ARTag is attached in the upper part of the sword that the user is holding. Baxter uses his head camera to detect the ARTag in order to know the approximate pose of the user’s sword. We choose No.23 ARTag.
In our project, the size of ARTag is chosen to be 3.5 cm by 3.5 cm. Through testing, we found that the ARTag can be detected from a distance about 70 cm from the head camera, with a perspective of 150 degrees. This is within the reachable range of the robot. We chose this as the designed defending space.
We use a gripper button control example from the SDK of Rethink Robotics. This program can be run in parallel with other code. We can control the gripper any time we want thus it iss easy for us to pass the sword to the robot. The circular button is used to open the gripper and the dash button is used to close the gripper.
- Algorithm
The algorithms for defense, fightback and pulling out sword action will be illustrated separately.
For defense, after getting the position of the ARTag, we will set the goal position of the right gripper. Since we let the robot always hold the sword vertically, the goal position of the the right gripper can be set to be just 10 cm below the position of ARTag. That is, x_goal = x_artag, y_goal = y_artag, and z_goal = z_artag - 0.10.
For fight back, we first need to know whether the attacker’s sword is to the right or the left of the robot. If it is to the right, the robot will hit from the left. If it is to the left, the robot will hit from the right.
For fight back, we first need to know whether the attacker’s sword is to the right or to the left of the robot. If it is to the right, the robot will hit from the left. If it is to the left, the robot will hit from the right.
For the pulling out sword motion, we set several different stages of action, which are explained in one of the videos. The core of this part is to make Baxter switch the hand holding the swort from the left gripper to the right gripper and to pull out the sword in a straight line without breaking it. While the left gripper is holding the sword, the right gripper will reach the sword in the direction of the y-axis of the left gripper, and then it will pull out the sword in the direction of the x-axis of the left gripper. We use the function “compute_cartesian_path” to achieve this.
- Orientation Constraints
To make sure that the robot will not execute weird trajectories or make the sword collide with his own body, we add orientation constraints for the gripper during the path execution. Orientation is expressed in quaternions. During defense, the right gripper is always pointing straightly out to the left.
Because of the reason that the robot waits for a long time before it re-plans a path after failing to do so, the path planning is adjusted down to 1 second, which greatly reduces the response time of the robot’s movement.
When a human holds the sword, it is difficult to hold it perfectly still, and it will most likely tremble some. The robot will plan and execute a new path every time it detects a change in position of the ARTag. As a result, the robot’s arm will shake during the defense. To avoid shaking, the algorithm is designed to plan a new path only if the distance between the new detected position and the previous detected position exceeds the threshold we set. In our program, the threshold is set to be 5 cm.
Design Choices and Tradeoff
- - Detection Accuracy vs Similarity to Real Sword
The real sword used in kendo competition is very thin. We tried to make our model sword similar to the real sword. However, since we need to attach an ARTag on it and it needs to be large enough for detection, the model sword can’t be too thin. We choose 3.6 cm as a compromised solution.
- - Detection Accuracy vs Defense Range
The reachable range of Baxter is about 100 cm in front of him. We wanted the range of defense of the robot to be as large as possible. However, due to the resolution of the head camera and the size of the ARTag we decided on using, we limited the defense range to be 70 cm, to guarantee its functionality and reliability.
We hoped the robot could use different poses for defense, but this lead to more failures of path planning and sometimes the robot would make the sword collide with its body possibly even hitting people near it. In order to make it safe, we added constraints to the right gripper to make sure it is always pointing straight to the left while moving (as shown in the picture). In this way, the robot will always hold the sword vertically, which also makes the defense algorithm simpler. However, this trades off the flexibility of the robot.
Design Evaluation
Baxter is a low-cost robot which is unfortunately not very accurate. However, our project is not too dependant on a high accuracy so we coordinated the defense range and the size of ARTag. Therefore our design is able to meet the accuracy requirements.
Due to our reliable algorithms, along with orientation constraints and shaking elimination, the system is robust. Throughout a lot of tests that we made, it worked as expected for most of the time.
One downside of our system is that it is not efficient enough. Limited by the delay in ARTag detection and the time needed for the path planning and execution processes, the robot can only response to static attack. That is, a human needs to hold a sword in a certain position without moving. Then it takes the robot about 3 seconds to perform the defense action. This means we are quite far from real-word applications. For a more efficient Kendo robot, it should be able to respond to a moving sword, which would require a real-time control.
