List of Ablation Studies

[hjsuh94]

Check these off if we think we have a sufficient conclusion to any of these questions....

• Comparison between doing RRT on exact vs. bundled linearization.
• RRT with Euclidean vs. Mahalanobis metric on reachability.
• Tradeoff between small step and dense samples vs. big step and sparse samples.
• Do trajopt during extend vs. without.
• Contact sampling with and without.

Feel free to add more as we progress!

[hjsuh94]

Conclusion to exact vs. bundled linearization - if we use exact linearization to construct BB^T instead of the bundled one, it turns out many of these B matrices do not have volume before contact, leading to very poor performance in planning.

[hjsuh94]

Conclusion to RRT with Euclidean vs. Mahalanobis metric. It turns out that Euclid performs better (!!!). We would need to think more about why.

I think this is because when we randomly sample a goal to look for the closest node, the goal is quite far from the existing points in the tree, and the local metric is a poor indicator of whether or not this set is actually reachable.

On one hand, I'm disappointed. On the other hand, maybe tuning a global distance metric is not so different from tuning the Q matrix in trajopt....

[hjsuh94]

Conclusion to the third issue: it seems that what's hard about contact-based problems is not that we can't take longer steps, but the finger gaiting procedure which has to happen on finer steps. Applying the same input with longer horizon does not seem to enhance the performance of the algorithm.

[hjsuh94]

Fourth issue: let's avoid this at all costs. It takes too long

[hjsuh94]

Fifth issue: contact sampling certainly seems to give better manipulability ellipsoids!

[hjsuh94]