sim2sim.py为运行的mujoco脚本，前提是安装了mujoco
sim2sim_class.py被上述文件调用，里面需要先安装phc smpl等库，参考h2o那篇安装
g1_scene和g1_w_imu_3_balls是模型文件，用于mujoco的模型导入
ACCAD_g1_Male2MartialArtsPunches_c3d_E7uppercutleft_poses.pkl为重定向好的模型文件，来自AMASS库
25_04_06_18-19-05_g1_add_ref_vel_and_ref_delta_posmodel_20000.pt为训练好的模型
由于原作者的task_obs是考虑全局信息的，这里我将原来的task_obs函数观测作了以下更改：
# 只返回参考动作相对于参考动作root坐标系下的坐标：
def compute_imitation_observations_teleop_no_RootPos(root_rot ,ref_root_pos, ref_body_pos, ref_body_vel, body_pos, time_steps):
obs = []
B, J, _ = body_pos.shape #B为环境数 ，J为track_link_ids的数目 + head or hands
heading_inv_rot = torch_utils.calc_heading_quat_inv(root_rot)
# heading_rot = torch_utils.calc_heading_quat(root_rot)
heading_inv_rot_expand = heading_inv_rot.unsqueeze(-2).repeat((1, body_pos.shape[1], 1)).repeat_interleave(time_steps, 0)
# heading_rot_expand = heading_rot.unsqueeze(-2).repeat((1, body_pos.shape[1], 1)).repeat_interleave(time_steps, 0)
##### 参考位置相对于参考root坐标系下的值
delta_ref_body_pos = ref_body_pos.view(B, time_steps, J, 3) - ref_root_pos.view(B, 1, 1, 3) # preserves the body position
# make some changes to how futures are appended.
obs.append(delta_ref_body_pos.view(B, time_steps, -1)) # 1 * timestep * J * 3
obs.append(ref_body_vel.view(B, time_steps, -1)) # timestep * J * 3
obs = torch.cat(obs, dim=-1).view(B, -1)
return obs
对应于sim2sim.py文件的：
def compute_imitation_observations_teleop_max_no_RootPos(self, ref_root_pos, ref_body_pos, ref_body_vel, root_rot, time_steps):
obs = []
B = 1
J = len(self._track_bodies_extend_id)
# B, J, _ = body_pos.shape #B为环境数 ，J为track_link_ids的数目 + head or hands
# ================
heading_inv_rot = torch_utils.calc_heading_quat_inv(root_rot) #0
# heading_inv_ref_rot = torch_utils.calc_heading_quat_inv(ref_root_rot)
heading_inv_rot_expand = heading_inv_rot.unsqueeze(-2).repeat((1, J, 1)).repeat_interleave(time_steps, 0) # 1 3 4
# heading_inv_ref_rot_expand = heading_inv_ref_rot.unsqueeze(-2).repeat((1, J, 1)).repeat_interleave(time_steps, 0) # 1 3 4
# ================
# T_quat = self.quaternion_transform(heading_inv_rot, heading_inv_ref_rot) # 1 3 4
# T_quat = T_quat.unsqueeze(-2).repeat((1, J, 1)).repeat_interleave(time_steps, 0)
# ================
local_ref_body_pos = ref_body_pos.view(B, time_steps, J, 3) - ref_root_pos.view(B, 1, 1, 3) # preserves the body position
local_ref_body_pos_obs = torch_utils.my_quat_rotate(heading_inv_rot_expand.view(-1, 4), local_ref_body_pos.view(-1, 3))
local_ref_body_vel_obs = torch_utils.my_quat_rotate(heading_inv_rot_expand.view(-1, 4), ref_body_vel.view(-1, 3))
# ================
obs.append(local_ref_body_pos.view(B, time_steps, -1)) # 1 * timestep * J * 3
obs.append(ref_body_vel.view(B, time_steps, -1)) # timestep * J * 3
obs = torch.cat(obs, dim=-1).view(B, -1)
return obs
sim2sim_class.py里的：
def _load_motion(self):
# motion文件安
motion_path = f'{MUJOCO_TREE_ROOT_DIR}/motions/ACCAD_g1_Male2MartialArtsPunches_c3d_E7uppercutleft_poses.pkl'
# urdf文件安
skeleton_path = f'{MUJOCO_TREE_ROOT_DIR}/unitree_robots/g1_19dof/xml/g1_phc.xml'
同样sim2sim_class.py里：
class Sim2simCfg():
class sim_config:
mujoco_model_path = f'{MUJOCO_TREE_ROOT_DIR}/unitree_robots/g1_19dof/xml/g1_scene.xml'
在sim2sim.py文件下：
if __name__ == '__main__':
load_model = f'{MUJOCO_TREE_ROOT_DIR}/models/STUDENT/25_04_06_18-19-05_g1_add_ref_vel_and_ref_delta_posmodel_20000.pt'
policy = torch.jit.load(load_model)
cfg = Sim2simCfg()
cfg.sim_config.mujoco_model_path = f'{MUJOCO_TREE_ROOT_DIR}/unitree_robots/g1_19dof/xml/g1_scene.xml'
run_mujoco(policy, cfg)