Newton Native Data and Selection API

Newton Native Data and Selection API#

Mental model#

The Newton backend exposes its engine-owned Model, State, Control, and optional Contacts objects through isaaclab_newton.physics.NewtonManager. Their Warp arrays are the live engine data, rather than values pulled into a separate per-asset view buffer. The model owns structural and static arrays and labels; a state owns evolving simulation arrays; and a control owns actuation inputs. Contacts are optional and depend on the active solver and collision path.

Lifecycle prerequisite#

Access Newton data after the simulation context has initialized the physics backend and built its model. Treat the objects as invalid after a model rebuild and reacquire them after Isaac Lab has reinitialized the simulation. This API is intended for code that can own the necessary lifecycle and synchronization responsibilities.

Access live engine data#

Use the manager accessors to obtain the current Newton objects:

from isaaclab_newton.physics import NewtonManager

model = NewtonManager.get_model()
state = NewtonManager.get_state_0()
control = NewtonManager.get_control()
contacts = NewtonManager.get_contacts()

body_poses = state.body_q
joint_forces = control.joint_f

isaaclab_newton.physics.NewtonManager.get_model() can construct a visualization shadow model when PhysX is active. The write semantics in this guide apply only when Newton is the active, authoritative physics backend.

Reuse an Isaac Lab selection#

Isaac Lab’s Newton-backed assets expose the same generic newton.selection.ArticulationView selection helper. For example, reuse an articulation’s root selection instead of constructing the matching selection again:

robot = scene["robot"]
selection = robot.root_view
joint_positions = selection.get_dof_positions(state)

Newton uses this generic, label-based selection concept for Isaac Lab articulations, rigid objects, rigid-object collections, and cables. It is a selection helper over model indices, not per-asset typed storage.

Create a generic selection#

Code that owns a matching model can construct its own selection from a model and a label pattern:

from newton.selection import ArticulationView

selection = ArticulationView(
    model,
    pattern="/World/envs/env_*/Robot",
)

Discover data and labels#

Discover the available labels, counts, state fields, and selection operations at runtime instead of assuming a fixed set of arrays or methods:

print(model.articulation_label)
print(model.body_label)
print(model.joint_label)

model_counts = {
    name: getattr(model, name)
    for name in dir(model)
    if name.endswith("_count") and isinstance(getattr(model, name), int)
}
selection_shape = {
    "instances": selection.count,
    "joint_dofs_per_instance": selection.joint_dof_count,
    "links_per_instance": selection.link_count,
}
state_fields = sorted(name for name in dir(state) if not name.startswith("_"))
selection_methods = sorted(
    name
    for name in dir(selection)
    if name.startswith(("get_", "set_"))
)
print(model_counts)
print(selection_shape)
print(state_fields)
print(selection_methods)

Read and write through a selection#

Selections provide typed convenience methods as well as generic string-keyed get_attribute() and set_attribute() methods. The generic methods expose engine properties that do not have dedicated selection methods. Clone a selected value before modifying it when you want the write to remain explicit:

import warp as wp
from newton import ModelFlags
from isaaclab_newton.physics import NewtonManager

rolling_friction = wp.clone(
    selection.get_attribute("shape_material_mu_rolling", model)
)
# Modify rolling_friction with a Warp kernel before writing it back.
selection.set_attribute(
    "shape_material_mu_rolling",
    model,
    rolling_friction,
)
NewtonManager.add_model_change(ModelFlags.SHAPE_PROPERTIES)

The string names a Newton model attribute rather than an Isaac Lab field. This example uses rolling friction because it has no dedicated selection method. Notify the manager with the flag appropriate to the property changed; other model writes can require a different flag. State writes that change generalized coordinates instead require forward-kinematics synchronization through NewtonManager.invalidate_fk() and NewtonManager.forward().

Ownership, synchronization, and invalidation#

Direct writes bypass Isaac Lab caches and shape/order validation. Some Newton solvers swap current and next state buffers, so reacquire isaaclab_newton.physics.NewtonManager.get_state_0() when code needs the current authoritative state on a later step. A selection can survive state-buffer swaps because it describes model indices, but it must be recreated after a model rebuild. Solver-specific generalized- and maximal-coordinate conventions remain authoritative.

Authoritative references#