something about the state_desc

[Ailsa1994]

I want to know the specific meaning of the features of state_desc, where can I get it?

[kidzik]

We are working on describing the details, it'll appear here http://osim-rl.stanford.edu/docs/nips2018/observation/

[Frawak]

The observation is still bit erroneous. The state description (returned by the client.env_step) has only 2 dimensions for the center of mass. The documentation states that there are 3.

Moreover, the get_observation method of the ProstheticsEnv brings up some things: 1) The description:

    # y, vx, vy, ax, ay, rz, vrz, arz of pelvis (8 values)
    # x, y, vx, vy, ax, ay, rz, vrz, arz of head, torso, toes_l, toes_r, talus_l, talus_r (9*6 values)
    # rz, vrz, arz of ankle_l, ankle_r, back, hip_l, hip_r, knee_l, knee_r (7*3 values)
    # activation, fiber_len, fiber_vel for all muscles (3*18)
    # x, y, vx, vy, ax, ay ofg center of mass (6)
    # 8 + 9*6 + 8*3 + 3*18 + 6 = 146

• There are 158 values.
• You insert 8 joints in the equation instead of 7.
• Hip_l and Hip_r have 3 values/dimensions per [pos,vel,acc] the others have 3 which makes 5 3+2 9=33
• There are 19 muscles.
• 8 + 9 6 + 33 + 3 19 + 6 = 158

2) The dimensions:

            cur += state_desc["body_pos"][body_part][0:2]
            cur += state_desc["body_vel"][body_part][0:2]
            cur += state_desc["body_acc"][body_part][0:2]

Why are we neglecting an entire dimension - having 2 in a 3D scenario?

[kidzik]

Thanks for reporting. Yes, unfortunately, there is a bug in the center of mass. We will roll out the branch with the full observation of the center of mass. https://github.com/stanfordnmbl/osim-rl/tree/ver2.1

We will also remove the observation vector and keep only the dictionary to avoid confusion. Projection is only an example and there are many issues with that (such as redundant values)

[Ailsa1994]

So, when changed the center of mass in observation, we have a 160-D observation?

[Ailsa1994]

In osim.py, there gives a fixed value 158 when describe the observation space, but when changed the center of mass, there should be a 160-d observation, maybe they should be changed together.

[kidzik]

Sorry for the delayed answer. In the next release, we will skip the "projected" observation and only the dictionary will be available. Everyone can then can construct the observation vector on their own. We will provide some examples though (close to what we get now with project=True). All in all, this ambiguity brings more confusion than value, while the dictionary response seems more useful.