Inquiry about the detailed dimension and configuration of Policy Inputs

[kcli-wy]

Hi,

I am studying the play_amo.py script and would like to clarify the exact input dimensions for the policy_jit model to ensure correct deployment.

Based on the script, the policy takes obs_tensor and extra_hist as inputs. I have the following observations regarding the dimensions:

1. obs_tensor
Is the total dimension 1318?

• obs_prop (93 dims): Calculated as $3 \text{(ang vel)} + 3 \text{ (rpy)} + 2 \text{ (diff yaw)} + 23 \text{ (pos)} + 23 \text{ (vel)} + 23 \text{ (action)} + 2 \text{ (gait)} + 15 \text{ (adapter output)}$.
• obs_demo (17 dims): Includes arm DOFs (8), velocity commands $V_x, V_y$ (2), yaw command (1), torso orientation (3), and height (3).
• obs_priv (3 dims): priviledge state, I think it may be linear velocity that are not available in reality so they are 0.
• obs_hist (930 dims): Flattened obs_prop from the last 10 frames ($10 \times 93$).

2. extra_hist (Extra History)

• Dimension (2325 dims): Is this a fixed vector of $25 \times 93$ based on the extra_history_len of 25 and n_proprio of 93?

Question

1. Could you please confirm if these calculations are correct or if there are specific normalization/scaling factors (other than the ones applied to ang_vel and dof_vel in the code) that we should be aware of when deploying the policy to the real robot?
2. A real robot has 29 degrees of freedom, and even with a dexterous end-effector, if the policy only inputs the RPY values for the shoulder and elbow joints, will it operate normally?
3. Why does the height input of obs_demo require three dimensions with same values?
4. What does the tenth dimension of the variable obs_demo represent, and why is it always 0?
   Thank you for your great work!