Newton gear insertion

CONTRIBUTORS.md

@@ -75,6 +75,7 @@ Guidelines for modifications:
 * Chenyu Cao
 * Chenyu Yang
 * Connor Smith
+* Curie Park
 * CY (Chien-Ying) Chen
 * David Cao-Mueller
 * David Leon

docs/source/policy_deployment/02_gear_assembly/gear_assembly_policy.rst

@@ -734,7 +734,41 @@ First, launch the training with a small number of environments and visualization
                       --num_envs 4 \
                       --visualizer kit
 
-This will open the Isaac Sim viewer where you can observe the training process in real-time.
+    .. tab-item:: Flexiv Rizon 4s + Grav (Newton)
+
+        .. tab-set::
+
+           .. tab-item:: uv (Recommended)
+
+              .. code-block:: bash
+
+                  uv run isaaclab train --rl_library rsl_rl \
+                      --task IsaacContrib-Deploy-GearAssembly-Rizon4s-Grav \
+                      --num_envs 2 \
+                      --visualizer newton \
+                      presets=newton_hydroelastic
+
+           .. tab-item:: isaaclab.sh / isaaclab.bat
+
+              .. code-block:: bash
+
+                  ./isaaclab.sh train --rl_library rsl_rl \
+                      --task IsaacContrib-Deploy-GearAssembly-Rizon4s-Grav \
+                      --num_envs 2 \
+                      --visualizer newton \
+                      presets=newton_hydroelastic
+
+        Use ``presets=newton_sdf`` to compare point-SDF contacts. To train a task-space
+        policy, use the ``IsaacContrib-Deploy-GearAssembly-Rizon4s-Grav-Newton-IK`` task
+        with the same Newton preset.
+
+        The centered SDF assets can also be tested with PhysX by using
+        ``presets=physx_sdf`` on the ``IsaacContrib-Deploy-GearAssembly-Rizon4s-Grav``
+        task. The default PhysX preset retains the legacy Factory assets and frame conventions
+        for existing checkpoints.
+
+This opens the selected viewer (Isaac Sim for PhysX or Newton Viewer for Newton) so you can
+observe the training process in real time.
 
 .. figure:: ../../_static/policy_deployment/02_gear_assembly/sim_real_gear_assembly_train.jpg
     :align: center
@@ -820,10 +854,59 @@ Now launch the full training run with more parallel environments in headless mod
                       --num_envs 256 \
                       --video --video_length 200 --video_interval 76800
 
+    .. tab-item:: Flexiv Rizon 4s + Grav (Newton)
+
+        .. tab-set::
+
+           .. tab-item:: uv (Recommended)
+
+              .. code-block:: bash
+
+                  uv run isaaclab train --rl_library rsl_rl \
+                      --task IsaacContrib-Deploy-GearAssembly-Rizon4s-Grav \
+                      --num_envs 256 \
+                      --visualizer none \
+                      presets=newton_hydroelastic
+
+           .. tab-item:: isaaclab.sh / isaaclab.bat
+
+              .. code-block:: bash
+
+                  ./isaaclab.sh train --rl_library rsl_rl \
+                      --task IsaacContrib-Deploy-GearAssembly-Rizon4s-Grav \
+                      --num_envs 256 \
+                      --visualizer none \
+                      presets=newton_hydroelastic
+
+The Newton hydroelastic preset uses package-local concave SDF collision assets. Its defaults
+separate the expensive collision rate from contact integration: the outer physics and collision
+tick is 100 Hz, while 20 solver substeps produce a 2 kHz solver rate. A policy decimation of three
+gives a 33.3 Hz control rate. The preset also uses a 5 mm shape gap and a bounded triangle-pair
+buffer sized for 256 environments per GPU rank. The arm uses solver-native actuator gravity
+compensation so gravity remains enabled for the gears, and the Grav linkage is held by physical
+PD actuators; no action term rewrites the selected gear pose.
+
+The default triangle-pair capacity is 4,194,304 per GPU rank. If Newton reports a triangle-pair
+buffer overflow, reduce the number of environments per rank or increase the capacity before using
+the resulting policy; overflowed candidate contacts are discarded.
+
+For four-GPU training, launch 256 environments per rank (1,024 total):
+
+.. code-block:: bash
+
+    uv run python scripts/reinforcement_learning/train_multigpu.py \
+        --rl_library rsl_rl --num_gpus 4 \
+        --task IsaacContrib-Deploy-GearAssembly-Rizon4s-Grav \
+        --num_envs 256 --max_iterations 1000 \
+        --visualizer none \
+        presets=newton_hydroelastic
+
 **Command breakdown:**
 
-- ``--num_envs 256``: Runs 256 parallel environments for efficient training
-- ``--video_length 200``: Each video captures approximately one full episode (``episode_length_s / (sim.dt * decimation)`` = ``6.66 / (1/1000 * 33)`` ≈ 200 steps)
+- ``--num_envs 256``: Runs 256 parallel environments per trainer process
+- ``--video_length 200``: Captures approximately one PhysX episode
+  (``6.66 / (1/120 * 4)`` ≈ 200 policy steps); a Newton episode is approximately
+  222 policy steps (``6.66 / (0.01 * 3)``)
 - ``--video_interval 76800``: Records a video every 76,800 environment steps (~every 150 iterations), producing ~10 videos over full training
 
 Training typically takes ~12-24 hours for a robust insertion policy. The videos will be saved in the ``logs`` directory and can be reviewed to assess policy performance during training.

source/isaaclab_tasks/changelog.d/curiep-newton-gear-insertion.minor.rst

@@ -0,0 +1,15 @@
+Added
+^^^^^
+
+* Added Newton point-SDF and hydroelastic-SDF physics presets and collision
+  assets to the Flexiv Rizon 4s gear assembly task.
+* Added a PhysX SDF preset that uses the same centered collision assets for
+  cross-backend validation while preserving legacy PhysX defaults.
+* Added a Newton inverse-kinematics task variant for task-space policies.
+
+Fixed
+^^^^^
+
+* Fixed Newton gear reset, reward, and termination frames so selected gears
+  target their shafts and physical grasps are measured at the fingertip
+  midpoint without changing the PhysX defaults.

source/isaaclab_tasks/isaaclab_tasks/contrib/deploy/gear_assembly/config/rizon_4s/__init__.py

@@ -23,6 +23,17 @@
     },
 )
 
+# Flexiv Rizon 4s with Newton inverse-kinematics actions
+gym.register(
+    id="IsaacContrib-Deploy-GearAssembly-Rizon4s-Grav-Newton-IK",
+    entry_point="isaaclab.envs:ManagerBasedRLEnv",
+    disable_env_checker=True,
+    kwargs={
+        "env_cfg_entry_point": f"{__name__}.ik_newton_env_cfg:Rizon4sGearAssemblyIKNewtonEnvCfg",
+        "rsl_rl_cfg_entry_point": f"{agents.__name__}.rsl_rl_ppo_cfg:Rizon4sGearAssemblyRNNPPORunnerCfg",
+    },
+)
+
 # Flexiv Rizon 4s - ROS Inference
 gym.register(
     id="IsaacContrib-Deploy-GearAssembly-Rizon4s-Grav-ROS-Inference",

source/isaaclab_tasks/isaaclab_tasks/contrib/deploy/gear_assembly/config/rizon_4s/ik_newton_env_cfg.py

@@ -0,0 +1,48 @@
+# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+from isaaclab_newton.envs.mdp.actions.newton_ik_actions_cfg import NewtonInverseKinematicsActionCfg
+from isaaclab_newton.ik.newton_ik_objectives_cfg import NewtonIKJointLimitObjectiveCfg, NewtonIKPoseObjectiveCfg
+from isaaclab_newton.ik.newton_ik_solver_cfg import NewtonIKSolverCfg
+
+from isaaclab.utils.configclass import configclass
+
+from . import joint_pos_env_cfg
+
+
+@configclass
+class Rizon4sGearAssemblyIKNewtonEnvCfg(joint_pos_env_cfg.Rizon4sGearAssemblyEnvCfg):
+    """Gear-assembly with a Newton inverse-kinematics (task-space) action for the Rizon 4s arm.
+
+    Replaces the joint-space :class:`~isaaclab.envs.mdp.RelativeJointPositionActionCfg` with a
+    Newton-solved relative end-effector pose action, so the policy commands end-effector motion
+    directly. This provides a six-dimensional task-space alternative to the seven-dimensional
+    joint-space action. The gripper remains a fixed mimic mechanism (not commanded by the policy).
+
+    Note:
+        Newton IK consumes the replicated robot prototype, so this variant requires a Newton preset
+        (``presets=newton_mjwarp``, ``presets=newton_sdf``, or
+        ``presets=newton_hydroelastic``); it is not compatible with ``presets=physx``.
+    """
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        # Command the physical flange frame used by the real Flexiv Cartesian controller.
+        self.actions.arm_action = NewtonInverseKinematicsActionCfg(
+            asset_name="robot",
+            joint_names=["joint1", "joint2", "joint3", "joint4", "joint5", "joint6", "joint7"],
+            controller=NewtonIKSolverCfg(optimizer="lm", jacobian_mode="analytic", iterations=24),
+            clip={".*": (-0.5, 0.5)},
+            objectives=[
+                NewtonIKPoseObjectiveCfg(
+                    body_name="flange",
+                    command_type="pose",
+                    use_relative_mode=True,
+                    scale=0.025,
+                ),
+                NewtonIKJointLimitObjectiveCfg(weight=0.1),
+            ],
+        )