Physics Simulation in Isaac Sim

Purpose

Configure PhysicsScene and per-prim rigid bodies, collisions, materials, joint drives, solver selection, and physics sensors with worked examples.

Prerequisites

• Built Isaac Sim ($ISAAC_SIM_DIR or _build/linux-x86_64/release).
• NVIDIA GPU with a current driver (nvidia-smi).
• Shell env contract from isaac-sim-orchestrator: $ISAAC_SIM_DIR, $ISAAC_LAB_DIR, $WORKSPACE_DIR.

Limitations

• Targets Isaac Sim 6 / Kit 110 unless a section states otherwise.
• Does not replace official NVIDIA documentation for unsupported edge cases.

Troubleshooting

Error / symptom	Cause	Solution
Extension or import not found	Wrong $ISAAC_SIM_DIR or stale build	Point env vars at _build/linux-x86_64/release or rebuild
Black or empty frames	Missing lights or non-RTX render mode	Add dome/key light; confirm RTX / PathTracing settings
Hang on stage load or first render	MDL compile or oversized stage	Follow isolation steps in isaac-sim-troubleshooting

Targets Isaac Sim 6.0+ / Kit 110. Both backends share UsdPhysics.*; backend-specific behavior is called out per section.

Available Scripts

Script	Purpose	Arguments
scripts/prim_physics_setup.py	Per-prim physics setup helpers for Isaac Sim / USD (Kit 110)	see script --help

Running scripts

From agent runtimes that expose skill execution helpers, invoke helpers with run_script():

run_script("scripts/prim_physics_setup.py", args=["--help"])

From a built Isaac Sim tree, run the same file with ./python.sh (Linux) or python.bat (Windows) from _build/*/release, or execute shell helpers directly when they do not require the simulator.

Backend selection (Kit 110)

isaacsim.core.simulation_manager registers physics engines and picks the active one. default_engine in its extension.toml is "physx", but the isaacsim.physics.newton extension defaults auto_switch_on_startup = true, so any app that enables isaacsim.physics.newton (the standard isaacsim.exp.full.kit does) ends up with Newton active at startup.

from isaacsim.core.simulation_manager import SimulationManager

SimulationManager.switch_physics_engine("newton")   # or "physx"
print(SimulationManager.get_active_physics_engine())

# Inspect what is available
from isaacsim.physics.newton import get_available_physics_engines, get_active_physics_engine
print(get_available_physics_engines())

Force the engine explicitly when launching:

--/exts/isaacsim.core.simulation_manager/default_engine=newton     # or =physx
--/exts/isaacsim.physics.newton/auto_switch_on_startup=false        # opt out of auto-switch

Newton config classes (extension Python API; not surfaced in the user-guide RST yet — see the extension's API docs page or docs/isaacsim/physics/newton_physics.rst):

Class	Role
isaacsim.physics.newton.NewtonConfig	per-sim settings (CUDA graph capture, fabric sync, contact/joint defaults)
XPBDSolverConfig	XPBD solver (rigid + soft)
MuJoCoSolverConfig	MuJoCo Warp solver
isaacsim.physics.newton.tensors	NumPy / PyTorch / Warp frontends

Both Newton and PhysX consume the standard UsdPhysics.Scene + PhysxSchema.PhysxSceneAPI; many PhysxSchema.* attributes are still honored under Newton, plus Newton reads its solver config via omni.usd.schema.newton.

Stack & reading order

1. This skill: scene config, per-prim setup, contact materials, drives, sensors, readback, backend selection.
2. usd-articulation: multi-link articulations + Robot Schema overlay.
3. urdf-mjcf-to-usd-conversion: importer config (RL vs teleop drives).
4. isaac-sim-troubleshooting: when physics misbehaves on Kit 110.

Mechanism recipes (impact, feeders, dominoes, tops, cradles, pendulum waves, escapements) live in Worked Examples.

---

Part 1 — Scene-Level Configuration

PhysicsScene Setup

from pxr import Usd, UsdGeom, UsdPhysics, PhysxSchema, Gf

ps = UsdPhysics.Scene.Define(stage, "/World/PhysicsScene")
ps.CreateGravityDirectionAttr().Set(Gf.Vec3f(0, 0, -1))
ps.CreateGravityMagnitudeAttr().Set(9.81)

px = PhysxSchema.PhysxSceneAPI.Apply(ps.GetPrim())
px.CreateTimeStepsPerSecondAttr().Set(240)   # see Hz table below
px.CreateEnableCCDAttr().Set(True)
px.CreateEnableStabilizationAttr().Set(True)
px.CreateSolverTypeAttr().Set("TGS")          # TGS or PGS; TGS preferred for articulations

Physics Hz Selection

Scenario	Hz	Notes
Standard rigid-body scenes	60–120	Default for warehouse, general sim
Stacking / contact-rich	240	Tight contact resolution
High-velocity impacts	120 with 2–4 substeps	Pair with CCD
Small-part vibration (feeders)	≥ 4× vibration freq, typically 480	Resolve oscillation correctly
Spinning bodies / gyros	480	Numerical precision for angular momentum
Stiff contact chains (cradles, escapements)	480	Solver needs many sub-iterations

Rule of thumb: physics timestep must be > 4× the highest frequency in the system (vibration, spin, contact-stiffness mode).

Solver Iteration Counts (per-body)

Set on PhysxRigidBodyAPI per body that needs it. Higher = more accurate, slower.

Scenario	Position iters	Velocity iters
Simple rigid bodies, tumbling	16	4
Stacking	32	8
Complex joints / articulations	64	16
Stiff contact chains (cradle, escapement)	64	32

pxrb = PhysxSchema.PhysxRigidBodyAPI.Apply(prim)
pxrb.CreateSolverPositionIterationCountAttr().Set(32)
pxrb.CreateSolverVelocityIterationCountAttr().Set(8)
pxrb.CreateEnableCCDAttr().Set(True)

When to Disable Stabilization

EnableStabilizationAttr is on by default and helps stacks settle. It destroys angular momentum on free-spinning bodies. Disable it for:

• Spinning tops, gyros, flywheels
• Pendulum mechanisms (clock escapements, pendulum waves)
• Anything whose correctness depends on conserved angular velocity

px.CreateEnableStabilizationAttr().Set(False)

---

Part 2 — Per-Prim Physics Setup

RigidBody / Collision / Static / Kinematic

Use scripts/prim_physics_setup.py for dynamic, static, and kinematic body setup. It keeps the RigidBodyAPI, MassAPI, and CollisionAPI application sequence in one executable implementation.

Rule: RigidBodyAPI + CollisionAPI on the same prim. Splitting them across parent/child causes intermittent collision failures.

Static Colliders with Scale — Translate-First Pattern

Scaling a Cube prim with CollisionAPI applied directly causes PhysX to use the wrong collision bounds (objects fall through ground). Use a parent xform for position, a child mesh for scale:

# CORRECT
xf = UsdGeom.Xform.Define(stage, "/World/Ground")
UsdGeom.Xformable(xf.GetPrim()).AddTranslateOp().Set(Gf.Vec3d(0, 0, -0.05))
mesh = UsdGeom.Cube.Define(stage, "/World/Ground/Mesh")
mesh.CreateSizeAttr().Set(1.0)
UsdGeom.Xformable(mesh.GetPrim()).AddScaleOp().Set(Gf.Vec3f(50.0, 50.0, 0.1))
UsdPhysics.CollisionAPI.Apply(mesh.GetPrim())

Cube.size=1.0 means the cube has half-extents of 0.5, not 1.0. Use size=2.0 when you want "the scale op equals the half-extent."

Mass & Inertia

mass_api = UsdPhysics.MassAPI.Apply(prim)
mass_api.CreateMassAttr().Set(0.25)                                    # kg
mass_api.CreateCenterOfMassAttr().Set(Gf.Vec3f(0, 0, 0.05))            # local
mass_api.CreateDiagonalInertiaAttr().Set(Gf.Vec3f(1e-4, 1e-4, 2e-4))   # kg·m²

For URDF-imported robots, prefer import_inertia_tensor: true in config.yaml over auto-computed geometric inertia (see urdf-mjcf-to-usd-conversion).

---

Part 3 — Contact Materials

def create_contact_material(stage, mat_path, static_friction=0.5,
                            dynamic_friction=0.4, restitution=0.1):
    prim = stage.DefinePrim(mat_path)
    mat = UsdPhysics.MaterialAPI.Apply(prim)
    mat.CreateStaticFrictionAttr().Set(static_friction)
    mat.CreateDynamicFrictionAttr().Set(dynamic_friction)
    mat.CreateRestitutionAttr().Set(restitution)
    return mat

Reference Values

Material pairing	Static μ	Dynamic μ	Restitution
Concrete on concrete	0.6	0.5	0.05
Steel on steel	0.74	0.57	0.6
Rubber on rubber	0.8	0.7	0.5
Rubber on concrete	1.0	0.8	0.3
Wood on wood	0.5	0.3	0.2
Metal generic	0.4	0.3	0.2
Plastic (dice)	0.4	0.3	0.3
Felt (casino)	0.5	0.4	0.2
Cardboard on steel	0.4	0.3	0.1

For chains of stiff contacts (Newton's cradle, escapements), set restitutionCombineMode=max on PhysxMaterialAPI so the highest restitution wins at each contact.

---

Part 4 — Joint Drives

joint = stage.GetPrimAtPath("/World/Robot/joint_arm")
drive = UsdPhysics.DriveAPI.Apply(joint, "angular")    # "angular" | "linear"
drive.CreateTypeAttr().Set("force")                    # "force" | "acceleration"
drive.CreateStiffnessAttr().Set(1000.0)                # Kp (Nm/rad for angular)
drive.CreateDampingAttr().Set(100.0)                   # Kd (Nm·s/rad)
drive.CreateMaxForceAttr().Set(500.0)                  # torque/force limit
drive.CreateTargetPositionAttr().Set(0.0)              # target (deg or m)

For RL training, the agent commands torques directly. Set drive_type to none and stiffness/damping to 0 in config.yaml (see urdf-mjcf-to-usd-conversion). Active PD drives fight the RL agent.

For revolute pendulum joints (clock escapements, pendulum waves), set joint friction to 0:

joint_api = PhysxSchema.PhysxJointAPI.Apply(joint)
joint_api.CreateJointFrictionAttr().Set(0.0)

---

Part 5 — Backend Selection (Newton vs PhysX)

Quick Choice

You want	Use
RL training with thousands of envs	Newton (Featherstone or MuJoCo)
Differentiable simulation	Newton
Legacy PhysX scene from Isaac Sim 5.x	PhysX
Soft bodies, cloth, deformables	Newton (VBD or XPBD)
Validated against MuJoCo baselines	Newton SolverMuJoCo

Newton Solvers

Solver	Coordinates	Differentiable	Best For
SolverFeatherstone	Generalized	Yes (Warp)	Articulated robots (default for manipulators, legged)
SolverMuJoCo	Generalized	Yes (mujoco-warp)	Validated locomotion, MuJoCo policy ports
SolverXPBD	Maximal	Partial	Soft constraints, cables, ropes
SolverSemiImplicit	Maximal	Yes (Warp)	Fast prototyping, simple rigid bodies
SolverVBD	(deformable)	Yes	Soft bodies, deformables

Newton vs PhysX Differences

Aspect	PhysX	Newton
Backend	Closed C++/CUDA	Warp/CUDA (open, JIT)
Coordinates	Maximal (6DoF per body)	Generalized (Featherstone) or maximal
Differentiable	No	Yes (native Warp autodiff)
Multi-GPU	Limited	Yes (Warp device abstraction)
USD integration	Schema extensions	Native USD loader
Performance ceiling	Good < 4096 envs	Designed for 10K+ envs

Newton + Torch — Critical Init Order

Never import torch before Newton physics settles — CUDA context conflict hangs Kit. Defer torch imports until after timeline.play() + settle loop. Use map_location="cpu" for policy inference if VRAM is tight.

Newton-Specific Configuration (Isaac Lab)

# config.yaml for URDF→USD conversion (Isaac Lab)
make_instanceable: true    # CRITICAL for RL parallel envs
fix_base: false            # true for fixed-base arm; false for mobile/legged

See urdf-mjcf-to-usd-conversion for the full schema.

---

Part 6 — Physics Sensors

The current namespace is isaacsim.sensors.experimental.physics (authoring + runtime classes paired). The legacy isaacsim.sensors.physics import path still works but is deprecated for new code.

Migration: see Migrating from isaacsim.sensors.physics to isaacsim.sensors.experimental.physics for the concept mapping and updated scripts.

Contact

from isaacsim.sensors.experimental.physics import Contact, ContactSensor
import isaacsim.core.experimental.utils.app as app_utils

contact = Contact.create(
    path="/World/Robot/foot/contact",
    min_threshold=0.0, max_threshold=1e6, radius=-1,   # -1 = use collision shape
)
sensor = ContactSensor(contact)
app_utils.play(commit=True)            # required before get_data()
reading = sensor.get_data()            # ContactSensorReading

IMU

from isaacsim.sensors.experimental.physics import IMU, IMUSensor

imu = IMU.create(path="/World/Robot/imu", tick_rate=200.0)
sensor = IMUSensor(imu, annotators=["linear_acceleration", "angular_velocity", "orientation"])
app_utils.play(commit=True)
frame = sensor.get_data()                # returns IMUSensorReading

Effort / joint state

Runtime-only classes from the same module; no separate authoring type — they wrap an existing joint by path.

from isaacsim.sensors.experimental.physics import EffortSensor, JointStateSensor

effort = EffortSensor("/World/Robot/joint_arm_1")
joint  = JointStateSensor("/World/Robot/joint_arm_1")
app_utils.play(commit=True)
reading = effort.get_data()              # EffortSensorReading
state   = joint.get_data()               # JointStateSensorReading

Raycast (scene query)

import omni.physics.tensors as physics_tensors

sim_view = physics_tensors.create_simulation_view("cuda:0")
hit = sim_view.raycast_closest(origin, direction, max_dist)

For higher-fidelity sensor simulation (LiDAR scan patterns, multi-ray, vendor sensor models, depth/radar/acoustic), see isaac-sim-sensor and isaac-camera. They use the modern isaacsim.sensors.experimental.rtx and .physics namespaces.

---

Part 7 — Physics-to-USD Readback (CRITICAL)

The most common silent bug: reading authored USD transforms instead of simulated state.

Source	Returns	When to use
UsdGeom.XformCache.GetLocalToWorldTransform()	Authored USD transform (initial pose)	Editor-time queries, before play
RigidPrim.get_world_pose()	Simulated state	Always during simulation
Articulation.get_world_poses() (Kit 110)	Simulated state for articulated bodies	Articulated robots
Dynamic Control (DC) dc.get_rigid_body_pose()	Simulated state	Legacy / quick scripts

Why XformCache Is Wrong During Sim

updateToUsd=True writes physics state to Fabric, not the USD stage layer. XformCache reads the USD layer. Result: it always returns initial poses.

RigidPrim / GeomPrim pattern (Kit 110)

from isaacsim.core.experimental.prims import RigidPrim
import isaacsim.core.experimental.utils.app as app_utils

rp = RigidPrim(paths="/World/Dice/Die_*")
app_utils.play(commit=True)
pos_wp, quat_wp = rp.get_world_poses()   # warp arrays
positions  = pos_wp.numpy()              # (N, 3)
quaternions = quat_wp.numpy()            # (N, 4) [w, x, y, z]

Articulation pattern (Kit 110)

from isaacsim.core.experimental.prims import Articulation
import isaacsim.core.experimental.utils.app as app_utils

robot = Articulation("/World/Robot")
app_utils.play(commit=True)        # required for tensor data
pos_wp, quat_wp = robot.get_world_poses()
dof_positions  = robot.get_dof_positions().numpy()    # (N, num_dofs)
dof_velocities = robot.get_dof_velocities().numpy()
J = robot.get_jacobian_matrices().numpy()             # for IK

Legacy isaacsim.core.api.articulations.Articulation / isaacsim.core.api.prims.RigidPrim still load but are superseded by the isaacsim.core.experimental.* stack.

Migration: for the broader omni.isaac.* → isaacsim.* renaming map, see Renaming Extensions. For the experimental Articulation / RigidPrim APIs see the Python scripting index.

Quaternion Convention

USD/Isaac uses [w, x, y, z]; scipy uses [x, y, z, w]. Convert:

from scipy.spatial.transform import Rotation
r = Rotation.from_quat([quat[1], quat[2], quat[3], quat[0]])
euler = r.as_euler('xyz', degrees=True)

---

Part 8 — Common Gotchas

1. CollisionAPI alone = static collider; RigidBodyAPI + CollisionAPI = dynamic.
2. Same-prim requirement: both APIs must be on the same prim.
3. Kinematic bodies: use CreateKinematicEnabledAttr().Set(True), not enable/disable on RigidBodyAPI.
4. Cube.size=1.0 = half-extent 0.5. Use size=2.0 if you want scale ops to equal half-extents.
5. physics:velocity USD attributes are ignored by PhysX at runtime. Use RigidPrim.set_linear_velocities() after timeline.play().
6. physics:angularVelocity is in DEGREES/second, not rad/s. Convert with math.degrees().
7. SimulationContext.step(render=True) is the only reliable physics-with-render advance. app.update() does not sync physics.
8. Experimental sensors need app_utils.play(commit=True) (or timeline.play()) before get_data(); do not call initialize() from the legacy World flow.
9. get_rigid_body_state() does not exist in Isaac Sim 5.1+; use RigidPrim.get_world_poses() from isaacsim.core.experimental.prims.
10. PhysX cannot resolve sequential momentum transfer in same-island contact chains (see Worked Example 4: Newton's Cradle).
11. Tunneling at high spin rates: compound colliders fail above ~50 rad/s after 5–6s. Use simpler convex hulls or higher physics Hz.
12. Before controller diagnosis: verify a PhysicsScene, active timeline, and expected collision APIs are present.

---

Worked Examples (impact, vibratory feeder, gyro, cradle, escapement)

See examples.md for details.