Am I doing this right? (Quadruped sim)

I wrote all my notes in a blog post in the hopes it gets more visibility.

https://www.marginallyclever.com/2022/07/friday-facts-16-updog-unityml/

Unity's Discord ML channel seems to be really dead. Did I miss something?

 

Hi - looking at your behavior parameters screenshot, I'd say the stacking value is way too high. You're feeding 2208 (69 x 32) input values to a 3 x 128 unit network! What exactly is your agent observing? If it's joint rotation angles, a stacking value of 2 or 3 at the most should suffice for inferring motion. You might not even need stacking, if the agent observes joint velocities.

 

@mbaske thanks! It certainly trains faster that way!

I'm trying a process of incremental changes to approach the right network design that will solve the problem. Is there a better way?

 

@mbaske Based on your feedback, I've eliminated stacking.

whole project is https://github.com/MarginallyClever/DogML/

My inputs are:
- 12 vector3s for position of each joint relative to the ArticulationBody root using

Code (CSharp):

1. Torso.transform.InverseTransformPoint(limits[i].obj.transform.position);

- 12 floats joint positions (getJointPositions)
- 12 floats joint velocities (GetJointVelocities)
- 12 floats joint accelerations (GetJointAccelerations)
- 1 bool for each foot (if contact with floor). I'd rather have amount of pressure against floor but IDK how yet.
- 1 float = height off floor / maximum standing height
- 1 vector3 which way is up

Code (CSharp):

1. Torso.transform.InverseTransformDirection(new Vector3(0,1,0));

- 1 vector3 body velocity normalized

Code (CSharp):

1. Torso.transform.InverseTransformDirection(Torso.velocity);

- 1 float body velocity clamped(0,10) and then / 10.
- 1 vector3 local angular velocity

Code (CSharp):

1. Torso.transform.InverseTransformDirection(Torso.angularVelocity).normalized * Torso.angularVelocity.magnitude;

My outputs are desired angles for each joint via SetDriveTargets. I do not mess with joint forces the way the Crawler and Walker demos do. I *think* ArticulationBody uses JointDrives for each axis ArticulationBody.xDrive.forceLimit is a different name from JointDrive.maximumForce used in Crawler/Walker so... not certain.

Any ideas?

 

Hi - I usually base my config on the example projects. In this case, the crawler seems to be the best fit. And then, yes, I would tweak things incrementally. Observations look good.