This is a GitHub repo for the submitted paper "A Self-Evolving Digital Twin System Architecture for Indoor UAV Management based on Online Deep Learning" to IROS 2024. In this repo, we will release the realization of the proposed method called online LSTM, collected UAV data, and other comparison methods.
| Models | Model setting |
|---|---|
| MLP | An input layer with dimensions [11, 50], two hidden layers with dimensions [50, 50], and one output layer with dimensions [50, 1]. The Adam optimization method is employed with a learning rate of 0.001. |
| LSTM | An LSTM layer with dimensions [11, 50] and a linear layer with dimensions [50, 1]. The Adam optimization method is employed with a learning rate of 0.001. |
| Transformer | An input embedding layer with dimensions [11, 500], a positional encoding layer with dimensions 400, two transformer encoding layers (400 dimensions, 10 attention heads, and a dropout probability of 0.1), and a linear decoder layer with dimensions [400, 1]. The AdamW optimization method is employed with a learning rate of 0.005. |
| Online MLP | The neural network is aligned with the MLP method. |
| OSELM | The random layer comprises 70 hidden units, employing the "sigmoid" activation function. |
| SRPRegressor | The number of members in the ensemble is 20. |
| OXTRegressor | The number of trees in the ensemble is 20. |
| Online LSTM | The neural network is aligned with LSTM method. |