- Link to (complete!) set of experiments with some documentation.
- Replace references to supplementary material with link to website/git.
- I think this is done? Feel free to correct me if wrong.
- Clarify the inputs & outputs of Algorithm 1.
- I'm not sure what requires clarification here exactly. We do list inputs/outputs. Is there something else we should add?
- Section 4 (experiments) overall needs revision for clarity
- Talked offline: This is more of a style thing on John's part, and he's not unhappy with the current version so it's not a hard requirement.
- Replace Tedrake website reference with an academic paper ref.
- Cite “use of higher order numerical methods to solve the adjoint sensitivity equation” in McAsey, Michael, Libin Mou, and Weimin Han. "Convergence of the forward-backward sweep method in optimal control." Computational Optimization and Applications 53.1 (2012): 207-226.
- Cite previous work on “discussion of the instabilities of neural ODEs” in Gholami, Amir, Kurt Keutzer, and George Biros. "ANODE: unconditionally accurate memory-efficient gradients for neural ODEs." Proceedings of the 28th International Joint Conference on Artificial Intelligence. AAAI Press, 2019.
- In second line of abstract, "possible construct" should be "possible to construct"
- It is not clear what the red line in Fig. 3 is, namely, what is the L2 norm of Backsolve error. Please give a mathematical description.
- The experimental setup of LQR is not clear. Please describe, for instance, the dimension of state and control, and system matrices
$A$ and$B$ . - Fix the spacing in Eq. (11) in Appendix A.
- What are the control policy implemented in differential drive robot and DiffTaichi Electric? I don't find the description.
- In third line of Appendix B, remove "a" before
$K$ .