Regarding the reward function and weights

[Serissa]

Table 1 in the paper lists the weights of the reward function. May I ask if these are the weights used during training？

[dwei-k]

Yes, these weights are used during training. We adjust the ratio between the locomotion reward and the style reward depending on the situation. While keeping the locomotion reward constant, the style reward typically ranges from 2 to 5.

[Serissa]

The paper mentions that the robot control is composed of 16 joint angles. What is the reason for not using 29 joints？

[dwei-k]

The proposed framework is also applicable to the 23 dof setting; however, it requires additional adjustments to the reward function as well as domain randomization, such as constraints on the waist and shoulder degrees of freedom. We chose the 16 dof configuration because it is sufficient to capture most gait patterns while allowing for a relatively straightforward validation of the proposed framework.

[criszhenhi-hue]

I would like to ask how effective is the training of an asymmetric architecture that directly uses deep image input in one stage? Will it be difficult to converge?