Understanding mirror observations in symmetric environment

[v-r-a]

Hello @rohanpsingh, thank you for this wonderful library!

I am a researcher looking to generate locomotion on stairs for a biped robot.

Could you explain a bit about how the mirror observation works? I could not understand the following lines:

LearningHumanoidWalking/envs/jvrc/jvrc_walk.py

Line 74 in e517f53

# define indices for action and obs mirror fns

base_mir_obs = [-0.1, 1, # root orient -2, 3, -4, # root ang vel 11, -12, -13, 14, -15, 16, # motor pos [1] 5, -6, -7, 8, -9, 10, # motor pos [2] 23, -24, -25, 26, -27, 28, # motor vel [1] 17, -18, -19, 20, -21, 22, # motor vel [2] ] append_obs = [(len(base_mir_obs)+i) for i in range(3)] self.robot.clock_inds = append_obs[0:2] self.robot.mirrored_obs = np.array(base_mir_obs + append_obs, copy=True).tolist() self.robot.mirrored_acts = [6, -7, -8, 9, -10, 11, 0.1, -1, -2, 3, -4, 5,] 

Why is there a -0.1? Also, not able to understand the other indices.

I passed --no-mirror and went ahead at the moment.

[v-r-a]

Got it.

Mirrored about the Sagittal plane. Roll and Yaw axes get flipped.

-0.1 is a tick to flip the 0th DOF.

PS:
The order of joints at the hip is Hip roll --- Hip pitch --- Hip yaw (not yaw-roll-pitch)
The order of joints at the ankle is ankle roll --- ankle pitch.