Torch2Bullet transition with system identification.

jc-bao / policy-adaptation-survey

This repository is for comparing the prevailing adaptive control method in both control and learning communities.

Apache License 2.0

7 stars 1 forks source link

Open jc-bao opened 1 year ago

jc-bao commented 1 year ago

The lower-level attitude rate controller does not match in Bullet and Pytorch-based environments.

Here is a one-step difference between that two different environments.	Roll rate	Motor torque

Rope force	object pos

The torch version has a faster response compared with the bullet version. :arrow_forward: need to study the actual torque generated by the object.

jc-bao commented 1 year ago

At 1s timescale:

The gap comes from the incorrectly applied external torque in the pytorch.

jc-bao commented 1 year ago

In PyTorch simulator, compare the torque caused by object and applied torque (which turn out to be 1).

But after making the gain larger, there is no other effects. Need to be further investigated. :part_alternation_mark:

jc-bao commented 1 year ago

The problem is addressed at #19

Current comparison between two simulators: