system control explanation

[BeNavon]

it's particularly fascinating how they managed to control joint movements without reinforcement learning.
could someone briefly explain how joint control works in this simulation?
thanks!

[chrisalbertson]

You ask how a robot could possibly work without using RL? Let me ask you another question: Can you provide a link to an open-source robot controller comparable to “Champ” that does, in fact, use RL? I doubt such a thing exists. To answer your question, it works like this. A path in 3D space is computed, and then periodically (at about 50Hz) a point is taken from the path and used as a foot position, inverse kinematics is used to compute the joint angles that place the foot to this location. motors are commanded to those angles. The rest is “just details”. To learn more, you need to ask about each detail, one at a time, but as part of the question tell us what you were able to figure out and what puzzles you. As an “issue,” I think it is best to word it as “we need to document the theory of operation at a fairly high level.” However, making such a request is about the same as volunteering for a big job.

…

On Mar 9, 2025, at 9:18 AM, BeNavon ***@***.***> wrote: BeNavon created an issue (chvmp/champ#151) it's particularly fascinating how they managed to control joint movements without reinforcement learning. could someone briefly explain how joint control works in this simulation? thanks! — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread. <#151> <https://github.com/notifications/unsubscribe-auth/ABQKNRRJ3HDKCATSGCI2DJT2TRSWFAVCNFSM6AAAAABYUNFXWWVHI2DSMVQWIX3LMV43ASLTON2WKOZSHEYDKNJSGI2TKOA> BeNavon created an issue (chvmp/champ#151) <#151> it's particularly fascinating how they managed to control joint movements without reinforcement learning. could someone briefly explain how joint control works in this simulation? thanks! — Reply to this email directly, view it on GitHub <#151>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABQKNRRJ3HDKCATSGCI2DJT2TRSWFAVCNFSM6AAAAABYUNFXWWVHI2DSMVQWIX3LMV43ASLTON2WKOZSHEYDKNJSGI2TKOA>. You are receiving this because you are subscribed to this thread.

[Quothd]

There have been attempts in the past (https://github.com/OpenQuadruped/spot_mini_mini) and the gym packages. But these are a little outdated now.
There is also the issue of getting accurate feedback from the servos (actual angles and power consumption) is somewhat harder and requires extra hw (which also adds to the overall space and weight requirements)