Backends and Presets

An environment defines the robot, scene, sensors, and task. A backend runs the physics or rendering. A preset selects a named configuration, so you can switch backends or task modes without editing Python.

To select a physics preset, add it to your launch command:

uv run isaaclab train --rl_library rsl_rl \
   --task Isaac-Cartpole-Direct physics=newton_mjwarp

The task stays Isaac-Cartpole-Direct. The preset changes the configuration used to launch it.

What you can select

Tasks can expose three kinds of choices:

Environment
├── physics=...    Which physics configuration runs the simulation?
├── renderer=...   Which renderer produces camera data?
└── presets=...    Which task-specific mode or config bundle is used?

The asset, sensor, and scene APIs stay the same across supported backends. See Backend architecture for how Isaac Lab selects the implementation.

Physics, rendering, and visualization are separate choices. For example, a camera environment can use Newton physics with the Newton Warp renderer, while the visualizer is selected independently with --viz.

Find what a task supports

Check which presets a task supports with --help:

uv run isaaclab train --rl_library rsl_rl \
   --task Isaac-Cartpole-Camera-Direct --help

The help output groups names by physics=, renderer=, and presets=. To list the combinations each registered environment can launch, run:

uv run isaaclab list_envs --show_presets

Each line lists a combination that passes configuration checks for a headless launch; [default] marks the task default. Some combinations are incompatible: OvPhysX and OVRTX cannot share a process with Isaac Sim / Kit, and tasks may impose other restrictions. Use --keyword to narrow the search; checking every combination takes a few seconds per task.

A task listed only with (defaults) uses its registered default configuration and does not expose alternatives. Passing a name that a task does not list is unsupported and fails during configuration validation.

Choose a selector

Selector	Example	What it changes
physics=NAME	physics=newton_mjwarp	Selects a physics configuration, including its backend and solver.
renderer=NAME	renderer=newton_renderer	Selects a renderer configuration for tasks that produce camera data.
visualizer=NAME[,NAME,...]	visualizer=newton_gl	Selects interactive viewers; visualizer=none disables them.
presets=NAME[,NAME,...]	presets=rgb	Applies task-specific choices such as observation modes, camera layouts, or compatible configuration bundles.

Task scripts also accept --physics NAME, --renderer NAME, and --visualizer NAME (--viz NAME). These aliases select the same task presets and are consumed before runtime launch; --renderer never sets Kit’s rendering mode. Explicit Kit settings belong in --kit_args. Explicit selections override script defaults; conflicting selections are errors.

Selectors can be combined. This command chooses Newton with the MuJoCo-Warp solver, the Newton Warp renderer, and RGB observations:

uv run isaaclab train --rl_library rsl_rl \
   --task Isaac-Cartpole-Camera-Direct \
   physics=newton_mjwarp renderer=newton_renderer presets=rgb

Use values listed for the task. Incompatible physics, renderer, sensor, or observation settings fail configuration validation.

Common backend choices

The exact list depends on the environment, but these names follow shared conventions:

Physics presets

Name	Meaning
isaacsim_physx	Isaac Sim PhysX. Some tasks use this as their default.
physx	Automatic PhysX-family selection. Isaac Sim PhysX is used when the runtime needs Kit; a configured OvPhysX alternative can be used for fully kit-less runs.
newton_mjwarp	Newton physics with the MuJoCo-Warp solver.
newton_kamino	Newton physics with the Kamino solver. Available for selected tasks and compatible assets.
ovphysx	OvPhysX for supported tasks running without Kit.

Renderer presets

Name	Meaning
isaacsim_rtx	Isaac Sim RTX renderer.
rtx	Automatic RTX-family selection. Isaac Sim RTX is used when physics, visualization, livestreaming, or another runtime choice requires Kit; otherwise OVRTX is used for a fully kit-less run.
newton_renderer	Newton Warp renderer. This is the default for tasks that use the multi-backend renderer preset.
ovrtx	OVRTX renderer for runs without Kit.

Automatic selectors such as physics=physx and renderer=rtx are opt-in. Defaults are concrete so that running a task without selectors is predictable. A solver is not a separate backend: newton_mjwarp and newton_kamino both use Newton but configure different solvers.

Defaults, presets, and fine-tuning

A preset replaces the complete configuration section at its location; it does not merge fields from two alternatives. Isaac Lab resolves configuration in this order:

1. Apply each preset config’s default choice.

2. Apply global choices from presets=....

3. Apply a preset targeted at a specific path, such as env.sim.physics=newton_mjwarp.

4. Apply scalar Hydra overrides, such as env.sim.dt=0.002.

To change one value after selecting a preset:

uv run isaaclab train --rl_library rsl_rl \
   --task Isaac-Cartpole-Direct \
   physics=newton_mjwarp env.sim.dt=0.002

Use physics= and renderer= for backend choices. Use presets= for task modes or to update several matching sections together. A path selector replaces only the named section.

Important

Keep behavior-changing presets the same when loading a checkpoint. An observation preset can change tensor shapes, and a policy trained with one observation mode may not load with another.

How task authors expose choices

Task authors define alternatives with PresetCfg and choose one as the default. For a multi-backend task, the preset wrapper belongs in SimulationCfg:

from isaaclab.physics import PhysxAutoCfg
from isaaclab.sim import SimulationCfg
from isaaclab.utils import configclass
from isaaclab_newton.physics import MJWarpSolverCfg, NewtonCfg
from isaaclab_ov.physics import OvPhysxCfg
from isaaclab_physx.physics import PhysxCfg
from isaaclab_tasks.utils import PresetCfg


@configclass
class PhysicsCfg(PresetCfg):
    isaacsim_physx = PhysxCfg()
    ovphysx = OvPhysxCfg()
    physx = PhysxAutoCfg(
        isaacsim_physx=isaacsim_physx,
        ovphysx=ovphysx,
    )
    default = isaacsim_physx
    newton_mjwarp = NewtonCfg(solver_cfg=MJWarpSolverCfg())


@configclass
class MyEnvCfg:
    sim: SimulationCfg = SimulationCfg(physics=PhysicsCfg())

Keep backend-specific values inside named configurations whenever possible. This keeps task logic shared and makes every supported choice visible from the command line.

When backend selection must also change simulation-wide settings such as the time step, a physics preset may instead contain complete SimulationCfg alternatives. The physics= selector recognizes these bundles from their physics field and applies the complete matching simulation configuration.

Backend architecture

Factories dispatch an object to the active backend implementation at construction time, so code can use the same public API without importing backend-specific modules directly.

Factory dispatch

All factories inherit from FactoryBase. They select a backend key, then locate its implementation by module path:

1. FactoryBase._get_backend() returns the backend_name of the SimulationContext.backend_manager instance. Adding another physics backend requires setting this class attribute on its manager.

2. The factory module path determines the backend module path. For example, isaaclab.assets.articulation maps to isaaclab_physx.assets.articulation, isaaclab_newton.assets.articulation, or isaaclab_ov.assets.articulation. The OvPhysX backend key uses the shared isaaclab_ov integration package.

3. The factory lazily imports the backend module and caches the implementation class in a registry.

User code: Articulation(cfg)
    │
    ▼
FactoryBase.__new__()
    │
    ├─ _get_backend()       → "physx", "newton", or "ovphysx"
    │    (reads SimulationContext.backend_manager)
    │
    ├─ _get_module_name()   → "isaaclab_physx.assets.articulation"
    │    (OvPhysX maps to the shared isaaclab_ov package)
    │
    ├─ importlib.import_module()
    │    (lazy load — only on first use)
    │
    └─ Return backend-specific instance

Renderers and visualizers instead select their implementations through their configuration’s class_type. Their selection is independent of physics; renderer instances are shared through the simulation registry described below.

Physics manager lifecycle

Each backend implements PhysicsManager, the abstract base class that owns its simulation lifecycle. Implementations initialize their engine from a SimulationContext, update kinematics with forward(), advance simulation with step(), reset state with reset(), and release resources with close().

The manager exposes PhysicsEvent callbacks for cross-backend lifecycle work. MODEL_INIT occurs during scene construction, PHYSICS_READY after physics initialization, and STOP before native resources are replaced or shut down. The concrete close() implementation dispatches the STOP event.

SimulationContext keeps native resources and renderers in one registry. get_or_create_backend(backend_cfg) reuses a resource when the configuration and its concrete type match; otherwise it calls instantiate(backend_cfg). Optional keyword arguments supply constructor dependencies on a cache miss; they do not change the sharing key. BackendCfg defines resource settings and identity. RendererCfg extends it for renderers; PhysicsCfg selects the physics manager.

Finalize configurations before registration, then treat them and their nested values as read-only. Use a new configuration to request different settings.

close_backend(backend) closes the exact registered object after consumers release their bindings. It does not compare configurations. Resources described by BackendCfg must implement close(); plain construction configs hold Python-owned data and need no teardown method. A failed close keeps the entry for retry. During simulation teardown, physics first invalidates camera data, then material writers, renderers, visualizers, and remaining native resources close before the stage.

Image views use this same registry. ImageViewCfg compares by declaration identity so two perspective cameras with identical initial settings remain independent; windows and recorders explicitly share a declaration to reuse its image.

Managers and native renderers expose their borrowed resource through backend. For example, sim.backend_manager.backend.model accesses the finalized native model of a Newton simulation. Closing a renderer releases its bindings, not the shared native resource. Exposing native handles does not replace SDP transport.

Clone preparation resolves the contexts requested by consumer configurations before native initialization. Contexts derive from ReplicateContext; its default prepare(sim, routing) keeps the declared routes. Overrides can merge routes for a shared representation and acquire resources through the simulation registry. For example, the OV package can replace native OVPhysX and OVRTX routes with one OVStage route. Renderer constructors declare requirements; they must not select another consumer’s cloning path. Preparation also runs before reset for standalone consumers registered after scene construction. Consumers registered after physics is ready use the same preparation before their initialization.

The resulting clone contexts are registered separately as sim.clone_contexts[Context] = Context(...) before plan dispatch. They apply the plan but do not own native runtime resources. Stage population, clone execution, and physics/render attachment follow the preparation decision.

Newton has two resources with different lifetimes, not two interchangeable backends:

• ModelBuilder holds mutable construction data. Cloning populates it and sensors declare requirements before finalization. It remains available for hard reset. NewtonBuilderCfg is a plain construction cfg, not a BackendCfg; the builder needs no native close().

• NewtonBackend owns the finalized model and native buffers. Physics and render consumers borrow those handles. Closing it releases runtime allocations without closing the builder.

Both resources use the same registry:

builder_cfg = NewtonBuilderCfg(physics_cfg=sim.cfg.physics)
builder = sim.get_or_create_backend(builder_cfg)
# Clone/import populates this builder before model allocation.
model_cfg = NewtonBackendCfg(physics_cfg=sim.cfg.physics, device=sim.device)
backend = sim.get_or_create_backend(model_cfg)

Both configurations use the selected physics cfg; non-Newton physics selects a render-only representation. SimulationContext has no backend-specific cfg fields, and consumers do not access clone contexts. Consumers request body transforms and visual points directly through SDP. Queries share the native resource’s BVHs but keep each consumer’s captured work separate.

Portable asset and sensor interfaces

Assets and sensors use the same layering as the factories:

1. A base class in isaaclab defines the public contract, such as BaseArticulation or BaseContactSensor.

2. A factory class inherits from both FactoryBase and that base class.

3. Backend packages provide the supported implementations.

Sensor data classes use the same pattern, for example ContactSensorData(FactoryBase, BaseContactSensorData). Implementations expose ProxyArray values through public asset and sensor data properties. Each proxy wraps the underlying wp.array and provides explicit .warp access to that array and cached, zero-copy .torch access to a torch.Tensor view. Use those accessors when an API specifically requires one representation. Passing a ProxyArray to wp.to_torch() is supported only by a deprecated compatibility shim; new code should use proxy_array.torch. Backend-native and internal storage may still use raw Warp arrays. See Working with ProxyArray for usage and buffer lifetime guidance.

Portable renderer and scene-data interfaces

Rendering is selected independently from physics. Acquire implementations of the BaseRenderer contract through sim.get_or_create_backend(renderer_cfg). The RenderContext coordinates their rendering lifecycle through a filtered view of that registry, without a separate renderer cache or renderer ownership. It validates global settings and registration timing, initializes renderers after physics is ready, and coordinates stage preparation, scene updates, and material writers. See Renderers for renderer choices and usage.

Physics managers expose live simulation data through SceneDataBackend. The SceneDataProvider owned by the simulation context converts and remaps that data for backend-independent consumers:

physics manager -> SceneDataBackend -> SceneDataProvider -> renderer or visualizer

This boundary lets renderers and visualizers consume a common Warp-native data path without knowing which physics engine owns the state. See Scene Data Provider for the complete data-flow model.

Native engine access boundary

The portable interfaces define the stable API boundary. Advanced code can use each engine’s native low-level data API, but those APIs intentionally keep their own ownership and synchronization semantics. See the backend guides for PhysX typed views, Newton model and state arrays, and OvPhysX tensor bindings.

Design principles

• Lazy loading: Backend modules are imported only when first instantiated, keeping startup fast and avoiding dependencies on unused backends.

• Recognized keys plus convention: Once the core selector recognizes a backend key, module paths mirror the isaaclab.X.Y structure. OvPhysX maps to isaaclab_ov.X.Y; other recognized backends use isaaclab_<backend>.X.Y by default.

• Independent selection: Physics backend, renderer, and visualizer are selected independently.

• Explicit data interop: Public asset and sensor data properties return ProxyArray; its .warp and .torch accessors expose the required array representation without copying.

• Zero runtime overhead: Selection occurs at instantiation time; it does not add dispatch logic to the simulation hot path.

Where to go next

• Environments lists environments and their supported presets.

• Hydra Configuration System covers scalar overrides, preset authoring, conflict handling, and advanced configuration behavior.

• Physics Backends compares physics backend runtime requirements, solver families, and intended uses.

• Renderers explains renderer selection and implementation details.