Experimental Results of EDRL in the DMControl-GB Environment

Taking the core performance comparison table as an example, the records in the data table cover the performance of different algorithms under two levels of environmental error: one is the video_ easy mode that replaces dynamic backgrounds but retains ground shadows, and the other is the video_ hard mode that removes ground and shadow references and has severe distribution offset. The row labels of the data table represent specific combinations of algorithms and testing tasks, while the column labels detail key quantitative indicators such as environment name, interference mode, average round reward (Mean), and standard deviation (Std).

The core unit of measurement involved is the "single round comprehensive return score", and the effective value range is strictly standardized between 0 and 1000. In addition, the dataset also includes a detailed table of ablation experiment records for the internal mechanism of the EDRL algorithm, including independent feature decoupling, reconstruction, state transition loss function, and multidimensional performance data when replacing the CNN/CNN-LSTM backbone network.

In terms of error control in data, the inherent statistical error of data mainly comes from the randomness in the exploration process of reinforcement learning and the inherent variance of neural network initialization. To address this reasonable error, this dataset has been normalized by displaying the mean and standard deviation.