AVBD Franka Simulation Notes

Rigid & Robot Tuning Guide

For this project, the key changes that made the AVBD Franka simulation work were:

1. Fixed AVBD drive-vs-limit behavior

   The main Newton-side fix was Donglai’s AVBD resolve_drive_limit_mode() change.

   Before, if a joint was slightly outside or numerically on a limit, AVBD could switch to LIMIT_UPPER / LIMIT_LOWER and suppress the PD drive. For Franka, that meant the commanded target could not pull the joint back from the limit, so the arm looked stuck or failed to track.

   The fix is: if the joint is outside a limit but the drive target is back inside the valid range, use DRIVE, not LIMIT.

   Conceptually:

   if q > lim_upper:
       if has_drive and drive_target < lim_upper:
           mode = DRIVE
       else:
           mode = LIMIT_UPPER
   

   Same for the lower limit.

2. Used one unified VBD solve

   The Franka, rigid contents, and deformable bag are all integrated by the same SolverVBD, so contact is two-way coupled instead of having the robot treated as an external or kinematic driver.

3. Increased AVBD structural joint constraint stiffness

   Once Franka was running in AVBD, the key parameter change was not more PD stiffness. It was making the AVBD structural joint constraints stiff enough:

   rigid_joint_linear_ke = 1.0e6
   rigid_joint_angular_ke = 1.0e6
   rigid_joint_linear_kd = 1.0e3
   rigid_joint_angular_kd = 1.0e2
   

   These make the articulated rigid bodies stay coherent under contact load. PD gains only control target tracking; they do not make the AVBD joint constraints themselves rigid.

4. Adjusted the PD control setup

   I kept the arm PD gains high enough for tracking, but did not solve the softness by simply increasing PD stiffness. The working setup uses strong but stable joint drives:

   arm_drive_ke = (1.0e6, 1.0e6, 8.0e5, 8.0e5, 6.0e5, 6.0e5, 6.0e5)
   arm_drive_kd = (1.0e5, 1.0e5, 8.0e4, 8.0e4, 6.0e4, 6.0e4, 6.0e4)
   gripper_drive_ke = 1.0e6
   gripper_drive_kd = 1.0e5
   

   The PD drives make the robot follow IK targets; the rigid_joint_* params make the robot physically rigid enough under contact.

5. Made the gripper contact strong enough to pinch the bag

   The stable pinch came from high-friction, dense finger pads and a small closed gap:

   soft_contact_mu = 1.0
   rigid_contact_hard = True
   finger_pad_density = 1000.0
   finger_shape_mu = 1.0
   finger_shape_ke = 1.0e6
   gripper_closed_gap = 0.001
   

6. Used the provided bag asset with self-contact off

   For the OBJ-bag Franka example, the bag asset is loaded from the provided USDA mesh. Bag self-contact is disabled because it did not materially improve the result while adding cost.

7. Kept rigid body contact stiffness moderate

   For the rigid contents, overly stiff material/contact ke and kd can make the bodies unstable. If rigid bodies jitter or explode, reduce their ke and kd first rather than increasing solver stiffness everywhere.

8. Tune contact distance per rigid body with separate margins

   Different rigid contents may need different contact distances. Use separate margins per rigid body or shape so each object can be tuned independently instead of relying on one global margin for all contacts.

Soft Tuning Guide

1. Try to make it work with the KFC bag asset

   Start from the target bag asset instead of overfitting the solver to a simplified proxy. The real asset exposes the mesh quality, scale, and contact details that the final demo needs to handle.

2. Improve convergence instead of blindly increasing stiffness

   To get a stiffer-looking bag, the first goal should be better convergence. Very stiff material settings, roughly tri_ke / tri_ka > 1e5, are already in the stiff regime. Increasing them further can hurt convergence, reduce diagonal dominance, and make the simulation look worse.

3. Use smaller time steps

   Smaller time steps improve conditioning and diagonal dominance, so VBD can converge to the intended stiff behavior more reliably.

4. Use lower resolution when debugging

   Lower cloth resolution reduces DoFs and makes convergence easier. Once the behavior is stable, increase resolution carefully.

5. Improve mesh quality

   Better triangle quality improves conditioning. Bad or highly irregular triangles make the local VBD solves harder and can amplify instability.

6. Turn off self-contact when it does not matter

   Bag self-contact is expensive and may not materially change the result for this pickup demo. Turn it off unless the visual or physical behavior clearly requires it.

7. Keep edge bending modest for poor-quality geometry

   Edge bending should not be set too high on bad-quality meshes. High bending stiffness on irregular geometry can fight the solver and introduce artifacts instead of improving shape preservation.