[Question] Gripper “tunnels” through deformable glove during pre-grasp, then latches closed (no contact-based gating)

[TakumaTsukakoshi]

Question

---

Gripper “tunnels” through deformable glove during pre-grasp, then latches closed (no contact-based gating)

Component(s): Isaac Lab env / deformable contact | Nextage gripper drives | grasp-latch logic

left_wireframe.webm

left_color.webm

Videos (evidence)

Summary

During the scripted pre-grasp phase the Robotiq fingers approach the glove mouth, but instead of squeezing the fabric they interpenetrate (“tunnel”) through the deformable mesh.

This looks like a combination of:

1. Deformable–rigid contact resolution too weak (low contact budget / offsets / resolution),
2. Very stiff/damped gripper drives pushing through the cloth

Environment details (from code)

• Physics: dt=1/120, decimation=2 → control at 60 Hz
• Solver iters (articulation root):

NEXTAGE_CFG = ArticulationCfg(
    spawn=sim_utils.UsdFileCfg(
        usd_path=os.path.join(PARENT_DIR, "assets/nextage_description/urdf/nextage_edit.usd"),
        activate_contact_sensors=True,
        copy_from_source=True,
        rigid_props=sim_utils.RigidBodyPropertiesCfg(
            disable_gravity=False,
            max_depenetration_velocity=5.0,
        ),
        articulation_props=sim_utils.ArticulationRootPropertiesCfg(
            enabled_self_collisions=True,
            fix_root_link=True,
            solver_position_iteration_count=8,
            solver_velocity_iteration_count=0,
        ),

• PhysX GPU budgets reduced:

  bounce_threshold_velocity=0.2
  gpu_temp_buffer_capacity=2**19
  gpu_max_soft_body_contacts=2**17
  gpu_collision_stack_size=2**19

• Deformable glove:

object_cfg: DeformableObjectCfg = DeformableObjectCfg(
        prim_path="/World/envs/env_.*/Object",
        init_state=DeformableObjectCfg.InitialStateCfg(pos=default_object_pos, rot=[0.7071, 0.0, 0.7071, 0.0]),
        spawn=UsdFileCfg(
            usd_path=object_usd,
            copy_from_source=True,
            visible=True,
            scale=(1.0, 1.0, 1.0),

            deformable_props=DeformableBodyPropertiesCfg(
                deformable_enabled=True,
                kinematic_enabled=False,
                self_collision=True,
                simulation_hexahedral_resolution=10,   
                collision_simplification=True,
                collision_simplification_remeshing=True,
                collision_simplification_remeshing_resolution=30, 
                collision_simplification_target_triangle_count=0,
                collision_simplification_force_conforming=True,
                contact_offset=0.005,
                rest_offset=0.003,
            ),
            visual_material=sim_utils.PreviewSurfaceCfg(
            diffuse_color=(0.8, 0.2, 0.2),
            opacity=1.0,             
        ),   
        ),
        debug_vis=False,
    )

• Material:

 physics_material=DeformableBodyMaterialCfg(
            youngs_modulus=8.0e7,     
            poissons_ratio=0.48,      
            density=1000.0,          
            damping_scale=1.0,
            elasticity_damping=0.012,       
            dynamic_friction=0.6,  
            ),

• Gripper drives (both hands):

  velocity_limit=0.2
  effort_limit=200.0
  stiffness=2000.0
  damping=2000.0

  _grasp_min_aperture = 0.018
  left/right_ee_euclidean_distance < _grasp_min_aperture ⇒ grasp = True

• After _grasp == True, the code latches the current gripper joints and enters policy phase with the latched pose held:

  self._grip_latched_q = latched.clone()
  self._policy_enabled = True
  ...
  self._set_grip_q(q_cmd, self._grip_latched_q)  # keeps grip fixed

Steps to reproduce

1. Launch env with a deformable glove object as configured in object_cfg (hexahedral res=10).
2. Start an episode (policy disabled initially).
3. Observe pre-grasp: gripper closes toward glove mouth.
4. Fingers penetrate the glove mesh; _grasp flips True from EE distance, without tactile.
5. Policy phase begins; gripper stays closed, holding nothing.

Expected vs Actual

• Expected: Fingers contact cloth, compress it, do not pass through.
• Actual: Cloth interpenetration occurs; grasp latches from EE distance alone

If there is anything that needs further discussion, please let me know.
I would really appreciate your support, and I look forward to your reply.

[elle-miller]

+1 - this would be very helpful to solve :)

[RandomOakForest]

Thank you for posting this. Great work. Your issue is a classic simulation artifact: the robotic gripper fingers tunnel through the deformable glove mesh during pre-grasp—instead of making realistic contact, the fingers pass through and grasping triggers from kinematic distance, not any tactile or actual collision. This is nearly always due to a combination of under-resolved deformable–rigid collision, control signal stiffness, and possibly mesh/contact settings in the simulation.

You can address this tunneling effect by tuning several recommended PhysX/IsaacLab settings that directly impact deformable–rigid contact detection and physical realism. Here are key steps to consider. I'll move this post to our Discussions section for follow up.

1. Increase Solver Iterations and Contact Resolutions

• Increase solver_position_iteration_count from 8 (default) to a higher value such as 16 or 32 for the gripper’s articulation root. More position iterations will let the solver spend more effort resolving deep penetrations and maintaining constraints during fast/frequent rigid–deformable contacts.
• Increase dt resolution: If possible, decrease your physics timestep (e.g. use $dt = 1/240$ if your simulation can handle it) for finer contact stepping, reducing the chance of fast-moving gripper parts missing collisions.
• Increase deformable hexahedral resolution: Your simulation_hexahedral_resolution=10 is modest. Try raising this (e.g. 14–24), keeping in mind that mesh resolution heavily impacts both contact fidelity and compute cost.

2. Adjust Contact Offsets and Budgets

• Increase deformable contact_offset and rest_offset slightly, e.g. contact_offset=0.01, rest_offset=0.006. This can help generate contacts before meshes visually overlap.
• Raise PhysX GPU budget parameters (especially gpu_max_soft_body_contacts and related buffers) if simulation gets starved for contact pairs or resolves only a subset per step.
• Check for PhysX contact visualization or logs to make sure sufficient deformable–rigid contacts are being generated.

3. Loosen Gripper Drive Parameters

Your gripper drives are both very stiff (stiffness=2000.0, damping=2000.0, effort_limit=200.0). Such high values let the gripper ignore collision constraints:

• Reduce stiffness and damping, especially for the pre-grasp phase. Try stiffness=400–800, damping=200–400. Allow some spring response so the gripper will “wait” for collision response.
• Velocity limit: Stay at or below 0.1 for slower approach in pre-grasp, letting simulation step more smoothly through contact events.

4. Ensure Correct Contact Filtering

Make sure both the gripper (rigid) and glove (deformable) are set up for proper collision in the USD or simulation:

• Both must have self-collision enabled (as you do).
• Check the simulation’s contact layers/masks for these prims to ensure they are not filtered out.

5. Enable Deformable Contact Debugging

Enable contacts visualization in Isaac Sim or use the debug_vis=True flag (if supported), allowing you to see when/where the actual contacts are being sampled between finger and cloth mesh.

6. Switch to Contact-Based Grasp Trigger

Instead of using only EE distance to trigger _grasp == True, use the actual contact information from PhysX (contact sensors or force sensors) and latch only when contacts between the gripper and the glove are detected—see [Isaac Sim Contact Sensor documentation].

---

Example Code Edits

Increase solver iterations:

ArticulationRootPropertiesCfg(
    # ...
    solver_position_iteration_count=16,  # Try 16, tune up if needed
    solver_velocity_iteration_count=0,
)

Increase cloth contact offsets:

DeformableBodyPropertiesCfg(
    # ...
    contact_offset=0.01,
    rest_offset=0.006,
)

Reduce gripper finger drive stiffness:

stiffness=600.0,
damping=300.0,

Check contact-based grasp in the step loop. Assume access to contact reports for the gripper:

if has_contact(gripper_fingers, glove):
    self._grasp = True