Managing package and package dependencies — In the beginning we ran into many errors because we were trying to code from ROS, Sawyer, the camera, and our vacuum gripper to work all at once. Since much of this code was dependent on each other, we ran into many dependency issues. To mitigate this, we chained our ROS workspaces together to keep our files separated while still being able to access them together.

Incorrect order of launch files + processes slowing down testing — We got stuck many times because we started nodes/launch files/processes in the wrong order and would result in errors. We also had to shut down and restart every process in the event of a failure which took up a lot of time and slowed down our development efficiency. To fix this, we wrote a new launch file that correctly started processes in the correct sequence efficiently.

Inconsistent motion plan — We used the MoveIt library to calculate motion plans between different positions on the Sawyer arm. Unfortunately the plan calculated was not consistent and would sometimes result in our robot doing wild motions. This would often knock down our setup or give us inconsistent results. To mitigate this, we implemented checkpointing and interpolated our poses. Instead of one continuous motion, we had the arm move to intermediate positions to manage wild/sudden movements

AR tags not facing the camera — When using the AR tags to try to calculate the angle between the object/gripper or object/wall, we found that the gripper kept rotating and pushing the AR tag out of the camera’s field of view. One potential solution to mitigate this was to use 2 cameras. However with trial and error, we found that enforcing an orientation constraint on the Sawyer joint proved to be the most reliable solution.

2D Implementation: Our solution is done assuming that the gripper and object are always in a single plane. This works well when the desired object angle is within that plane, but breaks down when we want to think about three dimensions.

Steady State Dynamics: Currently, we wait for the system to enter a steady state with no oscillations and no movement before taking measurements or using our model to make predictions. Our model is not able to actually predict a full motion plan of an object as it is manipulated along various trajectories. Rather, our model’s objective is to calculate desired quantities of the system at steady state. In the future, we would like to extend our model to compute the desired quantities and model the system dynamics more completely. Potentially, discretizing our system and applying our lagrangian dynamics model would allow us to compute the system state in time-steps when the system is not at its steady state.