Tuning collision penetration

[palanc]

Thanks for all of the work on this project! I am trying to use the shadow hand to grasp an object in Genesis. I have made modifications to the ik_shadow_hand.py example, but am finding that there is significant penetration between the fingers and object when grasping. Specifically, in the video below, you can see that the pinkie severely penetrates the cube as the hand closes. Could you provide any guidance on how to tune the collisions such that there is less penetration / more realistic? (I have included the code used to generate this example below as well)

pinkie_collide.webm

import argparse

import math
import numpy as np
import torch
import genesis as gs


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("-v", "--vis", action="store_true", default=False)
    args = parser.parse_args()

    ########################## init ##########################
    gs.init(seed=0, precision="32", logging_level="debug")

    ########################## create a scene ##########################
    scene = gs.Scene(
        viewer_options=gs.options.ViewerOptions(
            camera_pos=(2.5, 0.0, 1.5),
            camera_lookat=(0.0, 0.0, 0.5),
            camera_fov=40,
        ),
        show_viewer=args.vis,
        rigid_options=gs.options.RigidOptions(
            gravity=(0, 0, -9.81),
            enable_collision=True,
            enable_self_collision=False,
            enable_joint_limit=True,
        ),
    )

    ########################## entities ##########################
    robot = scene.add_entity(
        morph=gs.morphs.URDF(
            scale=1.0,
            file="urdf/shadow_hand/shadow_hand.urdf",
            pos=(0.0, 0.0, 0.1),
            euler=(-90, 0, 0),
            fixed=True,
        ),
        surface=gs.surfaces.Reflective(color=(0.4, 0.4, 0.4)),
    )

    plane = scene.add_entity(
        gs.morphs.Plane(),
    )

    cube = scene.add_entity(
    gs.morphs.Box(
        size=(0.05, 0.05, 0.05),
        pos=(0.0, 0.32, 0.2),
    )
)

    ########################## build ##########################
    B = 5
    scene.build(n_envs=B, env_spacing=(0.5, 0.5))

    lower_limits, upper_limits = robot.get_dofs_limit()

    # set positional gains
    robot.set_dofs_kp(
        kp=np.array(lower_limits.shape[-1]*[5.0]),
    )
    # set velocity gains
    robot.set_dofs_kv(
        kv=np.array(lower_limits.shape[-1]*[0.5]),
    )  
    robot.set_dofs_force_range(
        np.array(lower_limits.shape[-1]*[-3.0]),
        np.array(lower_limits.shape[-1]*[3.0]),
    )    
    robot.set_dofs_armature(np.array(lower_limits.shape[-1]*[0.01]))
    #robot.set_dofs_kp(kp=torch.ones((B, lower_limits.shape[-1]), device=lower_limits.device)*4500)
    #robot.set_dofs_kd(kd=torch.ones((B, lower_limits.shape[-1]), device=lower_limits.device)*450)    

    scene.reset()
    
    for i in range(20000):

        pos = (0.5*math.sin(i/100)+0.5)*(upper_limits - lower_limits) + lower_limits
        pos[0] = 0.0
        pos[1] = 0.0
        pos = torch.tile(pos, (B,1))

        robot.control_dofs_position(pos)

        scene.step()

        if i > 0 and i % 500 == 0:
            scene.reset()


if __name__ == "__main__":
    main()

[palanc]

I ended up loading the hand with a .mjcf instead of .urdf, which allows me to tune the collision interactions via the solimp parameters.

[duburcqa]

A public API has beeen added recently to allow modifying the constraint solver parameters of any geometry during simulation. I can give you more information if you are still interested.

[palanc]

Hi @duburcqa - Awesome that would be great if you could share the details

[duburcqa]

If you have a rigid entity, let's say:

    robot = scene.add_entity(
        morph=gs.morphs.URDF(
            scale=1.0,
            file="urdf/shadow_hand/shadow_hand.urdf",
            pos=(0.0, 0.0, 0.1),
            euler=(-90, 0, 0),
            fixed=True,
        ),
        surface=gs.surfaces.Reflective(color=(0.4, 0.4, 0.4)),
    )

then you can access links, equalities, and geometries and set/get the constraint solver parameters using sol_params/set_sol_params getter/setter. As an example, here is a unit test checking that everything works fine at API-level (not physics).

[duburcqa]

If you could share an example script that tunes these parameters to make it work for the shadow hand, that would be awesome. I would add it to the existing set of examples in Genesis repo.

By the way, URDF file loading has been updated to be more consistent with Mujoco. Could you run again your original script on the main branch? I'm curious to known if it works at least a bit better out-of-the-box.