collecting demonstrations in the robosuite

[Nimingez]

After collecting demonstrations in the robosuite env, and then converting them using the conversion script provided, there seems to be an issue of action not being in the scale [-1, 1] .
I have collected the demonstrations using "collect_human_demonstrations.py".
then python conversion/convert_robosuite.py
get_dataset_info.py
total transitions: 27087
total trajectories: 35
traj length mean: 773.9142857142857
traj length std: 240.8114106727588
traj length min: 362
traj length max: 1490
action min: -7.500000000000007
action max: 7.5

==== Filter Keys ====
filter key train with 32 demos
filter key valid with 3 demos

==== Env Meta ====
{
"type": 1,
"env_name": "Wipe",
"env_version": "1.4.1",
"env_kwargs": {
"env_name": "Wipe",
"robots": "UR5e",
"controller_configs": {
"type": "OSC_POSE",
"input_max": 1,
"input_min": -1,
"output_max": [
0.05,
0.05,
0.05,
0.5,
0.5,
0.5
],
"output_min": [
-0.05,
-0.05,
-0.05,
-0.5,
-0.5,
-0.5
],
"kp": 150,
"damping_ratio": 1,
"impedance_mode": "fixed",
"kp_limits": [
0,
300
],
"damping_ratio_limits": [
0,
10
],
"position_limits": null,
"orientation_limits": null,
"uncouple_pos_ori": true,
"control_delta": true,
"interpolation": null,
"ramp_ratio": 0.2
}
}
}

==== Dataset Structure ====
episode demo_1 with 1129 transitions
key: actions with shape (1129, 6)
key: states with shape (1129, 13)

Traceback (most recent call last):
File "get_dataset_info.py", line 134, in
raise Exception("Dataset should have actions in [-1., 1.] but got bounds [{}, {}]".format(action_min, action_max))
Exception: Dataset should have actions in [-1., 1.] but got bounds [-7.500000000000007, 7.5]

the actions not scaled down automatically

[Dhanushvarma]

I think if you clip the actions in the script it will fix the issue, this was also suggested by the authors in one of the previous issues, if I am not wrong.

like so :
action = np.clip(action, -1, 1)

[Nimingez]

Have you successfully collected data and used training for robomimic？ @Dhanushvarma

[Dhanushvarma]

What I have noticed is that, the keyboard interface is binary, and training with "keyboard-interface" collected data results in poor training. I think to overcome this, the solution is to use a "SpaceMouse" or some alternative interface.

[Abhiram824]

Hi, @Nimingez did the action clipping suggestion solve the issue?