Lightweight Imitation Learning Algorithm with Error Recovery for Human Direction Correction

Existing imitation learning methods for human directional corrections may lead to learning incorrect behaviors due to erroneous artificial teaching, resulting in a significant increase in the required number of iterations and even non-convergence situations, which can affect the system's performance. Additionally, the high computational complexity makes it unsuitable for embedded real-time application scenarios. To address these two issues, this study proposes a lightweight imitation learning algorithm that pre-corrects human-directed corrections. This method utilizes a deep learning network trained on a small dataset to correct human directional corrections and designs a lower-dimensional cost function for imitation learning. The proposed approach is applied to the example of a drone passing through doorways. Through the construction of a simulation platform and conducting simulation verification, the results show that the algorithm incorporating the correction error detection mechanism achieves an accuracy of over 98% in discerning human corrections, reduces training time by 27.87% per iteration, and decreases the average number of rounds by approximately 40%. The results indicate that the algorithm, which combines correction detection based on deep learning and a low-dimensional cost function, improves the accuracy of algorithm iterations, reduces computational complexity, and enhances computational speed.