Conclusion

Performance Evaluation

For the static target, the system worked well. We were able to get an average of 5.53 cm from the center of the target over 15 trials, and the bullet always hit the target. This was at a fixed distance of 66 cm.

With our initial object detection algorithm, we were able to hit the static target at a distance of between 100-300 cm 90% of the time. Between a firing distance of 100 cm and 120 cm, the average distance from the center of the target was 6.06 cm. Between a firing distance of 160-180 cm, the average distance from the center of the target was 4.24 cm. Between a firing distance of 200-220 cm, the average distance from the center of the target was 9.97 cm.

For the moving target, we were able to meet our target speed criteria of 5cm/s. At this speed and at a distance of 91 cm, we were able to get the bullet on the target 80% of the time. Between a firing distance of 100 cm and 120 cm, the average distance from the center of the target was 3.65 cm. Between a firing distance of 160-180 cm, the average distance from the center of the target was 5.82 cm. Between a firing distance of 200-220 cm, the average distance from the center of the target was 7.68 cm.

Ideally, we would have improved our system to the point where it would be able to more consistently hit the static target at a distance of 305-457 cm and a moving target at a distance greater than 220 cm.

Difficulties Encountered

Eliminating false positives in target detection.

Several items in the background, such as e-stops, tables, shirts, etc, were sometimes incorrectly detected as the target. Also, the bright lighting conditions and the quality of the wrist camera that we were using for computer vision didn't make it any easier.

Accurately measuring distance and position using the wrist camera without AR tag(s).

Since we decided against utilizing AR tags to detect the target, we utilized geometric calculations which were initially a bit off from the actual distance measurement.

Consistency of gripper setup.

Placing the gun in the gripper in a way that would allow the gun to fire consistently and accurately took a lot of trial and error. Sometimes even though we would setup the gripper the same way as we had the previous day, the gun would still be slightly inaccurate, or the grippers would fail to press the trigger. We had to figure out how much to close the gripper to hold the gun in position without firing the gun, and a scheme to fire the gun when in the holding position (since just closing the grippers at that position wouldn't work).

Improvements/Goals

• Move position faster/dynamically change path by hacking MoveIt.
• Use a better (e.g. depth sensing) camera for more accurate target acquisition and tracking
  • One slight improvement we made post-demo was to cut out a large yellow circle the same size as the target to allow us to detect the target further back. Although it did work better at further distances than the standalone target, we would have liked to be able to reliably use a smaller target.
• Try to hit smaller/faster moving targets. Our original target was to hit a paper airplane.
• We had to use a hacky gripper setup, which led to inaccuracies when shooting. In order to prevent these inaccuracies, it might be best to create a 3D printed gun holder to ensure consistent shooting orientation
• More reliable continuous tracking (possibly with a custom path planner)