Backend Architecture

Backend Architecture#

Overview#

Isaac Lab supports multiple physics backends while presenting common asset, sensor, and scene interfaces to environment code. 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. For choosing a backend or preset in an environment, see Backends and Presets.

Factory dispatch#

All factories inherit from FactoryBase. They locate supported backend implementations through a core backend-key selector followed by package and module-path conventions:

  1. The name of SimulationContext.physics_manager is mapped to one of the backend keys recognized by FactoryBase._get_backend(). Adding another physics backend requires extending this core selector.

  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.physics_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 owns native resources and renderer instances in one registry. get_or_create_backend(backend_cfg) reuses one resource for equal configurations of the same concrete type; a cache miss constructs instantiate(backend_cfg). BackendCfg describes resource settings and identity, and RendererCfg extends it for renderer instances. PhysicsCfg selects a physics manager. Finalize configurations before registration and treat them, including nested values, as read-only afterward. Use a new configuration for different settings. close_backend(backend) closes the exact registered object after all consumers have released their bindings; it does not compare or hash configurations. Resources declared through BackendCfg must implement close(). Plain construction cfgs share Python-owned data without a teardown operation; removing the registry entry releases its reference. A failed close retains the entry for retry. After physics shutdown invalidates camera render data, simulation teardown closes material writers, renderer instances, visualizers, and remaining native resources, in that order, before closing the stage.

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

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.

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.

Data classes use the same pattern, for example ArticulationData(FactoryBase, BaseArticulationData). 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 Native Physics API Access for PhysX typed views, Newton live model/state arrays and generic selections, 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.