Hi again. I am messing around with the code for months. I did read the paper over and over again, seems like controller are for synthesizing motion.

So.. I have made my own version of environment to train the characters to compete each other. and I have notice there is num_parallel variable which visualize numerous characters.

However, as it is on the code, it seems that action space dimension is only for single character unlike other things as root_history which includes dimension of num_parallel.

Is num_parallel variable exists just for training? I am confused..

Hello, I have some qusetion about the target angle in mocap_env.py.

The defination of target angle is

target_angle = (
            torch.atan2(target_delta[:, 1], target_delta[:, 0]).unsqueeze(1)
            + self.root_facing
        )

which seems to be a relative angle plus current facing angle.

But in calculating the reward, you just calculate the cos of target angle instead of the cos of relative angle

def calc_progress_reward(self):
        _, target_angle = self.get_target_delta_and_angle()
        direction_reward = target_angle.cos().add(-1)
        speed = self.next_frame[:, [0, 1]].norm(dim=1, keepdim=True)
        speed_reward = (self.target_speed_buf - speed).abs().mul(-1)
        return (direction_reward + speed_reward).exp()

Did I understand something wrong?

Hello, thank you for the great work!

While running vae_motion/train_controller.py, An error shows up as below:

logger.warn( Traceback (most recent call last): File "train_controller.py", line 240, in main() File "train_controller.py", line 206, in main obs = env.reset(reset_indices) TypeError: reset() takes 1 positional argument but 2 were given

Can I ask how it would be fixed?

The paper mentiones that use the joints position and orientation at the same time to represent a motion. However, I think one of them is able to represent a character adequately. Also, I noticed that the code render the character after the function set_joint_positions. I think it use the positions of each joint, but I am not sure what role does each joint's orientation play. I would appreaciate it if you can reply!

Thanks for your great work. I use https://github.com/ubisoft/ubisoft-laforge-animation-dataset to train from sractch. the reading bvh data code is referred by

from fairmotion.data import bvh

motion = bvh.load(BVH_FILENAME)

positions = motion.positions(local=False)  # (frames, joints, 3)
velocities = positions[1:] - positions[:-1]
orientations = motion.rotations(local=False)[..., :, :2].reshape(-1, 22, _6)

I only adjust 22 to 21 because new. dataset has 21 joint. here is my last epoch logger 237m 45s (- 0m 00s) (140 100.0%) | Recon: 1.618e-02 | KL: 2.287e-07 | PP: 0.000e+00; And then, I play it by gym with your code in random work mode. the result is not normal, can you give me some advices.

Hi, I am reading your codes and have problem in evaluate_actions when updating ppo:

• https://github.com/electronicarts/character-motion-vaes/blob/main/algorithms/ppo.py#L95

I notice that you get dist_entropy along with action and value loss, which function in backward propagation. Though dist_entropy doesn't work in your code since the entropy_coef currently in your code is 0 as default, I am still curious about how it functions and why you use this (What exactly "An ugly hack for my KFAC implementation." is :stuck_out_tongue_closed_eyes:)

Thanks

Hello,

In the article corresponding to this repository, in section 7.4 Acyclic Motions, is is mentioned an idea about additionally conditioning the VAE model. It is said that it can generate the next frame considering a one-hot encoding of an acyclic action like kick or header.

I have gone through the repository code, but I might have missed this part. As far as I have seen, at least for the default model (Mixture VAE), the encoder only conditions on the given number of previous frames (https://github.com/electronicarts/character-motion-vaes/blob/main/vae_motion/models.py#L87). Did I miss this additional conditioning part or it was not included in current state of the repository?

Thank you!

Hi, How to generage mocap.npz, which seems not easy to me. Can u give a clue how to generate mocap.npz from public mocap dataset?

The train_mvae.py script assumes the mocap data to be at environments/mocap.npz. The original training data is not included in this repo; but can be easily extracted from other public datasets.

Thanks very much!

BEST