Some question about config to adapt to new robot

[phuongboi]

Thank you for your contribution. I have a custom robot that similar to unitree g1 23dof style. It weight about 40 kg, and similar height to g1. So can I expect that I modify the g1 config to adapt to my custom robot. Can I ask some questions about your config? I noted that pos of both g1 and h1 is setting at (0, 0, 0.5). But usually they set 0.8 for g1 and 1.0 for h1, my robot usually set at 0.85, so which number I should put to this setting. I also note that you ignore wrist joint in default joint angles, but take account 3 waist joints while you are using g1 23 dof which is lock 2 waist dof

[TaoHuang13]

Hi. Regarding initial height, 0.5m is suitable for all robots because the robot is initialized with lying down postures rather than standing up. Regarding the G1 23 DOF, it has only one waist joint and one wrist joint. I write wrist_roll joint in default joint angles. 'waist_roll' and 'waist_pitch' can be ignored because they are not supported by G1 23 dof.

[phuongboi]

Thank you for clarifying. I saw wrist_roll in default joint https://github.com/OpenRobotLab/HoST/blob/0c2902cd6bf2e67877eca53f90383f50743df0f1/legged_gym/legged_gym/envs/g1/g1_config_ground.py#L53.

[phuongboi]

Hi, I would like to ask one more question. I am training on my custom robot. After 2000 iteration, the robot have passed phase 1 in training. But it seem to can not move it legs. Maybe there are some mechanical design problems. May I can ask about how many iteration that g1 finished all 3 phases?

[TaoHuang13]

Hi, 2000 iter is enough for G1 to finish all phases. Have you tuned hyperparameters related to the height? They should be customized based on your robot.

[phuongboi]

Thank you for answering. My robot have similar height to g1 so I used the same config as yours. https://github.com/OpenRobotLab/HoST/blob/0c2902cd6bf2e67877eca53f90383f50743df0f1/legged_gym/legged_gym/envs/g1/g1_config_ground.py#L187 It is also 23 dof with same config as g1. I also add keyframe session in robot's urdf. I may try with "wall" environment. I hope this env will less challenge for my robot

[phuongboi]

Hi, This is my training log after 257 iteration, the mean reward get around -400. I noted that reward_dof_pos_limits may be the major factor. Do you see any abnormal in my training process? Thank you

[TaoHuang13]

Hi, dof pos limit seems consistent with G1.

Could you please share all curves visualized in tensorboard?

[phuongboi]

Hi, This is the tensorboard that I train with 2500 iteration

[TaoHuang13]

Got it! The dof_pos_limit seeds are weird. It should be improved during training. Could you please (1) screenshot more curves in 'Episode Cards' and (2) play the policy and record the video of the motion?

[phuongboi]

I add 3 more screenshot in episode card. My robot seem to spread hand and leg to reach the target pose, it can not standing but sitting on the edge of platform. I hope you can imagine my robot's pose(due to company's regulation, currently I can not share the picture of robot).

[phuongboi]

Hi, I did some modification like increase dof limit in urdf, increase kp, kd, increase base_height_target and target_base_height_phase1,2,3 (because the lower body of our robot is higher than g)
https://github.com/OpenRobotLab/HoST/blob/0c2902cd6bf2e67877eca53f90383f50743df0f1/legged_gym/legged_gym/envs/g1/g1_config_ground.py#L183. Finally, our robot can standing up but its legs seem to have weird pose like the image.