Scalable Deep Unfolding of Conic Optimizers
Makes deep unfolding work for large conic solvers via matrix-free implicit differentiation and a stable PSD-projection gradient; reports up to 50× speedups on SDP-based planning.通过无矩阵隐式微分和稳定的半正定投影梯度，让深度展开适用于大规模锥优化求解器；报告在基于 SDP 的规划上最高提速 50 倍。
BibTeX
@misc{oshin2026scalable,
title = {{Scalable Deep Unfolding of Conic Optimizers}},
author = {Alex Oshin and Rahul Vodeb Ghosh and Evangelos A. Theodorou},
year = {2026},
eprint = {2606.13825},
archivePrefix = {arXiv},
primaryClass = {math.OC},
url = {https://arxiv.org/abs/2606.13825},
}Planning, Control & Safety规划、控制与安全Robot Learning: RL & Imitation机器人学习：强化学习与模仿学习optimizationdeep-unfolding