The robotics community has been paying more attention to coverage functions due to their variant applications (e.g., spatial search and mapping, etc.). Due to their submodularity, greedy algorithms can find solutions with theoretical guarantees for maximizing coverage problems even if these problems are NP-hard. However, learning coverage functions is still a challenging problem since the number of function outcome for N sets is \(2^N\). Moreover, transfer learning of coverage functions is unexplored. This research focuses on the transfer learning of coverage functions via utilizing the invariant properties in the Fourier domain. The proposed algorithms based on these properties can construct Fourier support for learning coverage functions. Experiments conducted with these algorithms show that the robot can learn the coverage functions using less samples than the prior learning approaches in different environments. Experiments also show that the lossless compression rate of the proposed algorithms is up to 40 billion.

由于覆盖函数的不同应用（例如，空间搜索和映射等），机器人社区一直更加关注覆盖函数。由于它们的子模块性，贪婪算法可以找到具有最大化覆盖问题的理论保证的解决方案，即使这些问题是 NP-hard。然而，学习覆盖函数仍然是一个具有挑战性的问题，因为N 个集合的函数结果的数量是\(2^N\). 此外，未探索覆盖函数的迁移学习。本研究侧重于通过利用傅立叶域中的不变性对覆盖函数进行迁移学习。基于这些特性提出的算法可以为学习覆盖函数构建傅立叶支持。使用这些算法进行的实验表明，在不同环境下，机器人可以使用比先前学习方法更少的样本来学习覆盖函数。实验还表明，所提算法的无损压缩率高达400亿次。