Mismatch in Observation Order Between go2 and Other Environments

[juanjuanjuanbing]

Thank you for your great work! I have encountered a problem when identifying observations. In the legged_robot.py environment class used in go2, the order of observations in compute_observations is as follows :
self.obs_buf torch.cat(( self.base_lin_vel * self.obs_scales.lin_vel, self.base_ang_vel * self.obs_scales.ang_vel, self.projected_gravity, self.commands[:, :3] * self.commands_scale, (self.dof_pos - self.default_dof_pos) * self.obs_scales.dof_pos, self.dof_vel * self.obs_scales.dof_vel, self.actions ), dim=-1)
However, in environments such as go1 and Mujoco deployments, the order of observations is:
self.obs_buf = torch.cat(( self.base_ang_vel * self.obs_scales.ang_vel, self.projected_gravity, self.commands[:, :3] * self.commands_scale, (self.dof_pos - self.default_dof_pos) * self.obs_scales.dof_pos, self.dof_vel * self.obs_scales.dof_vel, self.actions, sin_phase, cos_phase ), dim=-1)
If I want to train using go2, which order is correct? Theoretically, I should trust the former, but in my implemented network structure, I found that using the latter order for observations allows them to be correctly identified.

[craipy-hub]

I just provide a simple routine for your reference. If you want to train Go2, you can use any obs, but you just need to align them on different platforms.

[heartInsert]

@juanjuanjuanbing Did you solve this?