GCIMOPT
收藏资源简介:
GCIMOPT数据集由格勒诺布尔阿尔卑斯大学团队开发,包含通过FATROP求解器生成的多种控制任务最优轨迹。该数据集涵盖倒立摆稳定、平面/三维四旋翼稳定及6自由度机械臂点位控制等场景,每条轨迹包含状态-目标对及对应最优控制信号。通过基于中间状态目标重标记的数据增强技术,原始数据规模可扩展10倍。数据集专为训练轻量化(<8万参数)且实时性高(较优化求解器加速6000倍以上)的目标条件策略而设计,适用于资源受限控制器的部署。
The GCIMOPT dataset was developed by the research team from Université Grenoble Alpes. It encompasses optimal trajectories for a wide range of control tasks generated using the FATROP solver. This dataset covers scenarios such as inverted pendulum stabilization, planar/3D quadrotor stabilization, and 6-degree-of-freedom (6-DoF) robotic arm point-to-point control, among others. Each trajectory consists of a state-target pair and the corresponding optimal control signal. Through data augmentation based on intermediate state target relabeling, the size of the original dataset can be expanded by 10-fold. The dataset is specifically designed for training lightweight target-conditioned policies with fewer than 80,000 parameters that deliver high real-time performance, achieving a speedup of over 6000 times compared to conventional optimization solvers, and is suitable for deployment on resource-constrained controllers.

- 1GCImOpt: Learning efficient goal-conditioned policies by imitating optimal trajectories纳瓦拉公立大学·统计、数学与计算机科学系; 格勒诺布尔阿尔卑斯大学·国家综合理工学院 · 2026年



