This past week was spent again on integration, as well as on the final presentation slides. We hope to be done with integration in the next two days, leaving us some time do testing and benchmarking before the poster deadline.
Testing & Validation
The tests that are highest in priority have to do with correctness (finding a valid and optimal path) and performance (achieving some speedup). Path comparison is simpler, and is done via comparison with our correct C implementation of RRT + A*. Performance will be benchmarked via timing of the RRT kernel (and compared against the software version). We will also measure end-to-end performance to see how our improvements affect the system overall.
Perception system: Tested by comparing the dimensions of objects in a real-world scene and the mapped scene. This is >90% correct overall. However, the perception system wasn’t able to assign correct global coordinates to the objects (it was only able to preserve the relative distances and dimensions of the objects), so we made design changes to apply a fixed transformation to all points in the point cloud before processing the points using
octomap. This works and does not limit the scope of our project because the camera is not expected to move in real-world settings, so it is reasonable to measure and calibrate the angles of the 3d transformation when deploying the system.