OJ's Monte Carlo Localization

Introduction

Meet OJ, the Cozmo robot. OJ is a small robot that can be programmed to do a variety of tasks. Using Anki Cozmo we were tasked with improving an implemation of Monte Carlo Localiztion. The goal of the robot is to be able to solve the Kidnapped robot problem.

Goals

Develop a Robust Localization System: Create a reliable Monte Carlo Localization system that enables Cozmo to accurately determine its position within different environments, overcoming challenges such as uncertainty and dynamic changes.

Enhance Cozmo's Navigational Abilities: Improve Cozmo's ability to navigate back to its initial "home state" even when faced with disruptions or being moved to an unknown location, effectively solving the Kidnapped Robot Problem.

Future Improvements

• Currently, Cozmo determines how far away a block is from him based on the area of the square he draws around it. It would be useful to know approximately what sizes correspond with what distances in mm.
• As of now, Cozmo is only able to gather intelligence on where he starts. The next step would be to have him find his way back once moved.
  • To do this, our group talked about having him spin until he's seen all of the blocks, marking their distance from him.
  • Once he's done that, he should move towards one of the blocks according to some logic (some ideas we had were moving towards the furthest box, or trying to move in between the two closest blocks).
• While localizing his position, he should stop every few steps to recalibrate his direction.
• Once he has returned to what he thinks is the middle, he should use the localization method implemented by Ben Durham's group to ensure his proper direction.

Conclusions

• We had a great deal of confusion and difficulty when we tried to connect to the Cozmo SDK via Android while Apple seemed to connect with far less difficulty. A more detailed explanation of how to connect to the Cozmo SDK as well as an ios debugger might be more helpful (Python 3.7).
• Potential improvements / adjusting parameters seemed to vary in their effectiveness. For example, lowering the pose population to more sparsely sample the pose space seemed to improve the time by roughly 10 seconds as well as reduce the computational load. Conversely, other parameters such as number of pictures and angles before and after the localization process resulted in Cozmo just getting more confused (taking more time) and more likely to guess an incorrect home location.
• It also seems that since Monte Carlo Localization has been implemented successfully in prior years' code, we can potentially improve certain areas such as computational complexity, processing time, and localization accuracy. However, these goals of the project could also be expanded into different types of localization such as 2D localization to increase the complexity and creativity of the projects and methods developed rather than seeking to improve small aspects of the prior year's code. This would also allow for the Anki Cozmo Project to grow into branches of increasingly complex feats for the generations of CS371 students to come.