CS 188 Project

Project Goal: We implemented a DMP-based algorithm and neural network to solve the Robosuite Nut Assembly Task using expert demonstrations, then compared the training time and accuracy of each of the methods.

To do this to implement two separate methods for learning a policy from data:

1. DMPs: We implemented a Cartesian-space variant of dynamic motion primitives that is capable of generating both position and orientation trajectories.

2. A neural network capable of directly mapping sensor inputs to control outputs

Success Criteria

Criteria for a successful policy: Trained policy can successfully & consistently complete the Square Assembly task, including for targets with varying orientations.

Goal: Develop successful policies (i.e. meeting the above criteria) using separate DMP and deep learning approaches, and compare the results.

• Metrics for comparison: success rate, average time to complete the task

Between our two approaches, the DMP had a higher success rate, but took significantly longer to complete the task than the neural network.

We had a few guesses as to why:

• The neural network generally took very efficient, direct paths to target, whereas the DMP would often take winding, snaky, or otherwise inefficient paths (due to the imitated demonstration likely taking such a path originally).

• The runtime behavior of the DMP is significantly more parameter-dependent than the neural network; as such, the DMP is much more susceptible to having bad parameters/gains adversely affect performance, as was likely the case for our DMP.