[Question] Robotiq-85 grasping induces consistent end-effector rotation

[Gary-Yifan-Zhang]

When using the Robotiq-85 gripper (ImplicitActuatorCfg on finger_joint*) to grasp a box, the entire 6-DOF arm starts to rotate around a fixed world axis as soon as the fingers close.

• The phenomenon is deterministic (same direction every trial).
• Higher finger PD gains (stiffness ≥ 2e3, damping ≥ 1e2) make the rotation faster and more violent.
• Low gains (stiffness ≈ 1, damping ≈ 0.1) still produce drift, but slower.
• High finger gains amplify the left-right contact-force imbalance, driving the end-effector to rotate..
• The effect persists across different arm models (JAKA-ZU5, KuKa, E05 pro, etc.) and different Isaac Lab versions (4.0 → 4.5).

Here is a video of this problem,.

Screencast.from.2025.09.04.18.31.26.webm

Below is the configuration of my robotic arm.

# Copyright (c) 2022-2025, The Isaac Lab Project Developers.
# All rights reserved.
#
# SPDX-License-Identifier: BSD-3-Clause

"""Configuration for the E05 pro robots.

The following configurations are available:

* :obj:`E05_PRO_ROBOTIQ_CFG`: E05 pro robot with robotiq_85 hand
* :obj:`E05_PRO_ROBOTIQ_HIGH_PD_CFG`: E05 pro robot with norobotiq_85 hand with stiffer PD control

Reference:
"""

import isaaclab.sim as sim_utils
from isaaclab.actuators import ImplicitActuatorCfg
from isaaclab.assets.articulation import ArticulationCfg
from isaaclab.sensors import CameraCfg
from isaaclab.utils.assets import LOCAL_ASSET_ROOT_DIR, switch_robot_assets_path

##
# Configuration
##

robot = "E05_pro"
switch_robot_assets_path(robot)

E05_PRO_ROBOTIQ_CFG = ArticulationCfg(
    spawn=sim_utils.UsdFileCfg(
        usd_path=f"{LOCAL_ASSET_ROOT_DIR}/E05_pro_usd/E05_Pro_robotiq_85/E05_pro_robotiq_cam.usd",
        activate_contact_sensors=True,
        rigid_props=sim_utils.RigidBodyPropertiesCfg(
            disable_gravity=False,
            max_depenetration_velocity=5000.0,
            # solver_position_iteration_count=192,
            # solver_velocity_iteration_count=1,
            # max_contact_impulse=1e32,
        ),
        articulation_props=sim_utils.ArticulationRootPropertiesCfg(
            enabled_self_collisions=True,
            solver_position_iteration_count=16,
            solver_velocity_iteration_count=8,
        ),
        collision_props=sim_utils.CollisionPropertiesCfg(
            contact_offset=0.005,
            rest_offset=0.0,
        ),
    ),
    init_state=ArticulationCfg.InitialStateCfg(
        joint_pos={
            "joint_1": 1.571,
            "joint_2": -0.2618,  # -0.569,
            "joint_3": 1.571,
            "joint_4": 0.0,
            "joint_5": 1.3090,
            "joint_6": 1.571,  # 1.571
            "finger_joint": 0.0,
            "left_outer_finger_joint": 0.0,
            "right_outer_finger_joint": 0.0,
            "left_inner_finger_joint": 0.0,
            "right_inner_finger_joint": 0.0,
            "left_inner_finger_knuckle_joint": 0.0,
            "right_inner_finger_knuckle_joint": 0.0,
            # "right_outer_knuckle_joint": 0.0,
        },
    ),
    actuators={
        "e05_shoulder": ImplicitActuatorCfg(
            joint_names_expr=["joint_[1-4]"],
            effort_limit=870.0,
            velocity_limit=2.175,
            stiffness=80.0,
            damping=4.0,
        ),
        "e05_forearm": ImplicitActuatorCfg(
            joint_names_expr=["joint_[5-6]"],
            effort_limit=120.0,
            velocity_limit=2.61,
            stiffness=80.0,
            damping=4.0,
        ),
        "robotiq_hand": ImplicitActuatorCfg(
            joint_names_expr=["finger_joint*"],
            # joint_names_expr=["finger_joint*", "right_outer_knuckle_joint"],
            effort_limit=10.0,
            velocity_limit=3.0,
            stiffness=800,
            damping=4.0,
        ),
        "robotiq_inner_fingers": ImplicitActuatorCfg(
            joint_names_expr=[
                "right_inner_finger_joint",
                "left_inner_finger_joint",
            ],
            effort_limit=10.0, 
            velocity_limit=3.0,
            stiffness=800.0,            damping=4.0,
        ),
        "robotiq_inner_finger_knuckle": ImplicitActuatorCfg(
            joint_names_expr=[
                "right_inner_finger_knuckle_joint",
                "left_inner_finger_knuckle_joint",
            ],
            effort_limit=10.0,  
            velocity_limit=3.0,
            stiffness=800.0,
            damping=4.0,
        ),
        "robotiq_outer_fingers": ImplicitActuatorCfg(
            joint_names_expr=[
                "left_outer_finger_joint",
                "right_outer_finger_joint",
            ],
            effort_limit=10.0,
            velocity_limit=3.0,
            stiffness=800.0,
            damping=4.0,
        ),
    },
    # soft_joint_pos_limit_factor=1.0,
)
"""Configuration of E05 pro robot."""


E05_PRO_ROBOTIQ_HIGH_PD_CFG = E05_PRO_ROBOTIQ_CFG.copy()
E05_PRO_ROBOTIQ_HIGH_PD_CFG.spawn.rigid_props.disable_gravity = True
E05_PRO_ROBOTIQ_HIGH_PD_CFG.actuators["e05_shoulder"].stiffness = 800.0
E05_PRO_ROBOTIQ_HIGH_PD_CFG.actuators["e05_shoulder"].damping = 280.0
E05_PRO_ROBOTIQ_HIGH_PD_CFG.actuators["e05_forearm"].stiffness = 800.0
E05_PRO_ROBOTIQ_HIGH_PD_CFG.actuators["e05_forearm"].damping = 280.0
E05_PRO_ROBOTIQ_HIGH_PD_CFG.actuators["robotiq_hand"].stiffness = 2e3 # default 1.0
E05_PRO_ROBOTIQ_HIGH_PD_CFG.actuators["robotiq_hand"].damping = 1e2 # default 0.1
E05_PRO_ROBOTIQ_HIGH_PD_CFG.actuators["robotiq_inner_fingers"].stiffness = (
    0.002  # default 0.002
)
E05_PRO_ROBOTIQ_HIGH_PD_CFG.actuators["robotiq_inner_fingers"].damping = (
    0.00001  # default 0.00001
)
E05_PRO_ROBOTIQ_HIGH_PD_CFG.actuators["robotiq_inner_finger_knuckle"].stiffness = 0.0
E05_PRO_ROBOTIQ_HIGH_PD_CFG.actuators["robotiq_inner_finger_knuckle"].damping = 0.0
E05_PRO_ROBOTIQ_HIGH_PD_CFG.actuators["robotiq_outer_fingers"].stiffness = 1.0
E05_PRO_ROBOTIQ_HIGH_PD_CFG.actuators["robotiq_outer_fingers"].damping = 0.0

"""Configuration of E05 pro robot with stiffer PD control.

This configuration is useful for task-space control using differential IK.
"""

REALSENSE_CAMERA_CFG = CameraCfg(
    prim_path="{ENV_REGEX_NS}/Robot/E05_pro/camera_link/pinhole_camera",
    width=1280,
    height=720,
    spawn=sim_utils.PinholeCameraCfg(
        focal_length=24.0,
        focus_distance=400.0,
        horizontal_aperture=20.955,
        clipping_range=(0.01, 1.0e5),
    ),
    offset=CameraCfg.OffsetCfg(
        pos=(0.03534, 0.0, -0.09705), rot=(0.5, -0.5, -0.5, 0.5), convention="opengl"
    ),
    data_types=["rgb", "depth"],
)

I tried to add contact sensors on both fingers, here are the results for stiffness 1.0, damping 0.1:

tensor([[True]], device='cuda:0')
tensor([[[-0.4914, -6.7747,  0.0406]]], device='cuda:0')
tensor([[[ 0.4901,  6.7533, -0.0397]]], device='cuda:0')
tensor([[True]], device='cuda:0')
tensor([[True]], device='cuda:0')
tensor([[[-0.4914, -6.7747,  0.0406]]], device='cuda:0')
tensor([[[ 0.4901,  6.7533, -0.0397]]], device='cuda:0')
tensor([[True]], device='cuda:0')
tensor([[True]], device='cuda:0')
tensor([[[-0.4970, -6.8623,  0.0554]]], device='cuda:0')
tensor([[[ 0.4822,  6.6498, -0.0498]]], device='cuda:0')
tensor([[True]], device='cuda:0')
tensor([[True]], device='cuda:0')
tensor([[[-0.4970, -6.8623,  0.0554]]], device='cuda:0')
tensor([[[ 0.4822,  6.6498, -0.0498]]], device='cuda:0')
tensor([[True]], device='cuda:0')
tensor([[True]], device='cuda:0')
tensor([[[-0.4727, -6.6481,  0.0462]]], device='cuda:0')
tensor([[[ 0.4795,  6.7538, -0.0459]]], device='cuda:0')
tensor([[True]], device='cuda:0')
tensor([[True]], device='cuda:0')
tensor([[[-0.4727, -6.6481,  0.0462]]], device='cuda:0')
tensor([[[ 0.4795,  6.7538, -0.0459]]], device='cuda:0')
tensor([[True]], device='cuda:0')
tensor([[True]], device='cuda:0')
tensor([[[-0.4862, -6.7125,  0.0605]]], device='cuda:0')
tensor([[[ 0.4865,  6.7120, -0.0608]]], device='cuda:0')
tensor([[True]], device='cuda:0')
tensor([[True]], device='cuda:0')
tensor([[[-0.4862, -6.7125,  0.0605]]], device='cuda:0')
tensor([[[ 0.4865,  6.7120, -0.0608]]], device='cuda:0')

Here are the results for stiffness 2000, damping 100:

gripper_contact_l tensor([[[0., 0., 0.]]], device='cuda:0')
gripper_contact_r tensor([[[0., 0., 0.]]], device='cuda:0')
gripper_contact_l tensor([[[0., 0., 0.]]], device='cuda:0')
gripper_contact_r tensor([[[-1.2034e-03,  1.5110e+00,  1.5866e-05]]], device='cuda:0')
gripper_contact_l tensor([[[0., 0., 0.]]], device='cuda:0')
gripper_contact_r tensor([[[-1.2034e-03,  1.5110e+00,  1.5866e-05]]], device='cuda:0')
gripper_contact_l tensor([[[-0.2217, -5.8983, -0.0853]]], device='cuda:0')
gripper_contact_r tensor([[[0.1843, 4.8969, 0.0650]]], device='cuda:0')
gripper_contact_l tensor([[[-0.2217, -5.8983, -0.0853]]], device='cuda:0')
gripper_contact_r tensor([[[0.1843, 4.8969, 0.0650]]], device='cuda:0')
gripper_contact_l tensor([[[ -0.4000, -10.7384,  -0.1306]]], device='cuda:0')
gripper_contact_r tensor([[[0.3533, 9.4845, 0.1153]]], device='cuda:0')
gripper_contact_l tensor([[[ -0.4000, -10.7384,  -0.1306]]], device='cuda:0')
gripper_contact_r tensor([[[0.3533, 9.4845, 0.1153]]], device='cuda:0')
gripper_contact_l tensor([[[ -0.4327, -11.6632,  -0.1501]]], device='cuda:0')
gripper_contact_r tensor([[[ 0.4537, 12.2400,  0.1537]]], device='cuda:0')
gripper_contact_l tensor([[[ -0.4327, -11.6632,  -0.1501]]], device='cuda:0')
gripper_contact_r tensor([[[ 0.4537, 12.2400,  0.1537]]], device='cuda:0')
gripper_contact_l tensor([[[ -0.4314, -12.3871,  -0.2169]]], device='cuda:0')
gripper_contact_r tensor([[[ 0.5100, 14.6455, -4.4309]]], device='cuda:0')
gripper_contact_l tensor([[[ -0.4314, -12.3871,  -0.2169]]], device='cuda:0')
gripper_contact_r tensor([[[ 0.5100, 14.6455, -4.4309]]], device='cuda:0')
gripper_contact_l tensor([[[ -0.4584, -12.6838,  -0.2567]]], device='cuda:0')
gripper_contact_r tensor([[[ 0.5039, 14.3338, -6.1785]]], device='cuda:0')
gripper_contact_l tensor([[[ -0.4584, -12.6838,  -0.2567]]], device='cuda:0')
gripper_contact_r tensor([[[ 0.5039, 14.3338, -6.1785]]], device='cuda:0')
gripper_contact_l tensor([[[ -0.4323, -12.6140,  -0.2151]]], device='cuda:0')
gripper_contact_r tensor([[[ 0.5006, 14.9565, -4.4026]]], device='cuda:0')
gripper_contact_l tensor([[[ -0.4323, -12.6140,  -0.2151]]], device='cuda:0')

Thank you very much for your time and assistance. I truly appreciate your support and look forward to your valuable feedback.

[RandomOakForest]

Thank you for posting this. It is a great post for our Discussions section. I'll move this post there for follow up. Here is a summary to consider.

This effect is caused by imbalance in contact forces and actuation asymmetries when the Robotiq-85 gripper closes on an object, which, combined with high finger PD gains, introduces torques that the whole 6-DOF arm must compensate for—resulting in deterministic rotation around the fixed world axis.¹²

Underlying Mechanism

• The Robotiq-85 gripper uses two fingers actuated via finger joints, which ideally should close synchronously and exert equal and opposite forces on the object.²
• In practice, small mechanical or simulation asymmetries (even at the sub-resolution level) mean the fingers may not make perfect, simultaneous contact with the box. This results in unbalanced normal forces when contact is established, especially if the joints are commanded with high stiffness/damping (high PD gains).²
• Because the arm itself is only subject to virtual/environmental friction (much less than real-world bearings), these small imbalances are not damped out and rapidly accelerate the end-effector and the arm's base around the world axis.¹²

Amplification By Gains

• Higher finger stiffness/damping means larger corrective torques are generated for the same position error, amplifying any left/right contact force mismatch. This drives the rotation.²
• Low gains slow the actuator response, but the underlying unbalanced tangential forces still persist, creating slow drift instead of violent turning.²

Why the Effect Persists

• This is a simulation phenomenon caused by:
  • Floating-point error, interpolation, and solver tolerance on contacts.
  • Non-ideal actuator force distribution when the gripper fingers begin to contact the object off-center.
  • Lack of motor/sensor cross-coupling and no built-in anti-symmetric compensation.
• The effect is deterministic because the robot and environment initialize identically, so the same (small) imbalance occurs every run.¹²

Arm and Version Independence

• The phenomenon is independent of arm model and simulator version because it arises fundamentally from the physical simulation of two opposing actuators closing on a rigid object without sufficient real-world damping, compliance, or feedback correction.¹²

Mitigation/Tuning Notes

• Try tuning the gripper's Drive properties—including setting stiffness to zero and using high damping (e.g., 5000.0), relying on force (not position) control, and limiting the max applied force.²
• Also, careful alignment of the gripper's frame and object helps minimize these effects, as misalignment can exacerbate imbalances in contact.¹

Footnotes

1. https://github.com/isaac-sim/IsaacLab/discussions/2065 ↩ ↩² ↩³ ↩⁴ ↩⁵

2. https://docs.isaacsim.omniverse.nvidia.com/latest/robot_setup_tutorials/rig_closed_loop_structures.html ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

[hhaofan]

@Gary-Yifan-Zhang Thanks Gary! I have also faced this question and been troubled for a long time. That's a very nice question with as many as details. Great one! Also thanks @RandomOakForest for the detailed analysis!