diff --git a/source/isaaclab/changelog.d/task-matrix.rst b/source/isaaclab/changelog.d/task-matrix.rst
new file mode 100644
index 00000000000..6daa77bcefc
--- /dev/null
+++ b/source/isaaclab/changelog.d/task-matrix.rst
@@ -0,0 +1,9 @@
+Added
+^^^^^
+
+* Added ``tools/task_matrix.py``, which resolves whether a task's backend combination can
+  actually run. ``enumerate_task_presets`` reports what a task declares; ``resolve`` builds
+  the config and runs the runtime validator, so combinations that are declared but unusable
+  — ``isaacsim_physx`` with the kitless ``ovrtx`` renderer, for example — are identified
+  before a run is scheduled. Because it returns the resolved configs, automatic selectors
+  such as ``physics=physx`` are recognised by what they resolve to rather than by name.
diff --git a/tools/task_matrix.py b/tools/task_matrix.py
new file mode 100644
index 00000000000..3dc6dd8d12d
--- /dev/null
+++ b/tools/task_matrix.py
@@ -0,0 +1,278 @@
+# 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
+
+"""Resolve which backend combinations a task can actually run, and what they select.
+
+:func:`~isaaclab_tasks.utils.preset_cli.enumerate_task_presets` reports what a task
+*declares*, bucketed by selector target. It cannot say whether a declared combination
+**works**: OVRTX is kitless and cannot share a process with Kit physics, so
+``isaacsim_physx + ovrtx`` is declared yet unusable, and a task can advertise a backend
+it has no configuration for.
+
+:func:`resolve` answers that by building the config and running the runtime validator.
+It returns the configs the selection resolved to rather than a bare yes/no, which turns
+out to answer two further questions for free — both previously handled with
+hand-written name lists:
+
+* **Is a name an alias?** ``physics=physx`` resolves to :class:`PhysxAutoCfg`, which
+  picks OvPhysX kitless and Isaac Sim PhysX under Kit. An alias and the backend it
+  selects are the same run, so the environment tables list only concrete backends and a
+  dispatcher must not schedule both. :func:`is_selector` reads that off the resolved
+  config instead of comparing against the names ``physx`` and ``rtx``.
+* **Does a domain preset mirror a backend?** Some tasks give an untyped ``PresetCfg``
+  field a backend name, which lands under ``DOMAIN`` but is really selected through
+  ``physics=`` / ``renderer=``. A ``presets=`` entry resolving to a physics or renderer
+  config is such a mirror, whatever it is called.
+
+Deriving all three from one resolution keeps the answers correct as backends are added
+or renamed, and costs one config load per combination rather than a second pass.
+"""
+
+from __future__ import annotations
+
+from typing import Any, NamedTuple
+
+__all__ = ["Mode", "Resolution", "backend_names", "is_selector", "modes", "resolve", "resolves"]
+
+# Errors meaning the validator could not run at all, rather than that the combination
+# under test was rejected. Reporting these as "does not resolve" marks every
+# combination unusable and blames the caller's filters for an empty result.
+# ``TypeError`` is included because calling the validator with the wrong argument type
+# is otherwise indistinguishable from a rejected combination.
+_VALIDATOR_FAILURES = (AttributeError, NameError, SyntaxError, TypeError)
+
+
+class ResolutionError(RuntimeError):
+    """Raised when the runtime validator itself could not run."""
+
+
+class Mode(NamedTuple):
+    """One way to run a task.
+
+    Args:
+        physics: Physics preset token, or ``None`` for tasks that declare none.
+        renderer: Renderer preset token, or ``None`` to run headless.
+        presets: Domain preset token passed as ``presets=<value>``, or ``None``.
+            Exactly one at a time: presets targeting the same field conflict, so
+            ``presets=depth,rgb`` is rejected outright.
+    """
+
+    physics: str | None
+    renderer: str | None
+    presets: str | None
+
+
+class Resolution(NamedTuple):
+    """What a selection resolved to.
+
+    Args:
+        physics: Resolved physics config, or ``None`` when the task has none.
+        renderer: Resolved renderer configs, empty when the task runs headless.
+    """
+
+    physics: Any | None
+    renderer: tuple[Any, ...]
+
+
+def _renderer_cfgs(env_cfg: Any) -> tuple[Any, ...]:
+    """Return the renderer configs a resolved env config selected.
+
+    Physics sits at ``sim.physics``, but renderers attach to camera sensors and so can
+    be anywhere in the tree. Walking dataclass fields, dicts and sequences mirrors what
+    :func:`~isaaclab_tasks.utils.hydra.collect_presets` does for preset nodes.
+    """
+    import dataclasses
+
+    from isaaclab.renderers.renderer_cfg import RendererCfg
+
+    found: list[Any] = []
+    seen: set[int] = set()
+
+    def visit(obj: Any, depth: int = 0) -> None:
+        if obj is None or depth > 8 or id(obj) in seen:
+            return
+        seen.add(id(obj))
+        if isinstance(obj, RendererCfg):
+            found.append(obj)
+        if dataclasses.is_dataclass(obj):
+            for field in dataclasses.fields(obj):
+                visit(getattr(obj, field.name, None), depth + 1)
+        elif isinstance(obj, dict):
+            for value in obj.values():
+                visit(value, depth + 1)
+        elif isinstance(obj, (list, tuple)):
+            for value in obj:
+                visit(value, depth + 1)
+
+    visit(env_cfg)
+    return tuple(found)
+
+
+def resolve(
+    task_id: str, *, physics: str | None = None, renderer: str | None = None, presets: str | None = None
+) -> Resolution | None:
+    """Return what a preset selection resolves to, or ``None`` if it cannot run.
+
+    An unknown preset, an unloadable config, and a rejected backend pairing all mean
+    the same thing to a caller — the combination cannot run — so they return ``None``
+    alike. A task needing an extra that is not installed also returns ``None``: it
+    cannot run *here*, which is what the caller is asking.
+
+    Args:
+        task_id: Gym task id.
+        physics: Physics preset to select, or ``None`` to leave it unset.
+        renderer: Renderer preset to select, or ``None`` to leave it unset.
+        presets: Domain preset to select, or ``None`` to leave it unset.
+
+    Raises:
+        ResolutionError: If the validator could not run, e.g. because an Isaac Lab API
+            it depends on has changed. That is not the same as a rejected combination
+            and must not be reported as one.
+    """
+    import argparse
+    import sys
+
+    from isaaclab.app.sim_launcher import _get_kit_runtime_sources, _validate_runtime, scan
+
+    from isaaclab_tasks.utils import resolve_task_config, setup_preset_cli
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--task")
+    parser.add_argument("--agent", default=None)
+    argv = ["--task", task_id]
+    for token, value in (("physics", physics), ("renderer", renderer), ("presets", presets)):
+        if value is not None:
+            argv.append(f"{token}={value}")
+
+    original_argv = list(sys.argv)
+    try:
+        args, remaining = setup_preset_cli(parser, argv)
+        sys.argv = [sys.argv[0]] + remaining
+        env_cfg, _ = resolve_task_config(args.task, args.agent)
+        # Capture before scanning: ``scan`` collapses an automatic selector to the
+        # concrete backend it picks, which is exactly the distinction callers need.
+        selected = Resolution(
+            physics=getattr(getattr(env_cfg, "sim", None), "physics", None),
+            renderer=_renderer_cfgs(env_cfg),
+        )
+        # ``_validate_runtime`` takes the resolved Kit sources, not the parsed args.
+        # Passing args makes every scan look Kit-backed, which fires the OvPhysX guard
+        # for every OvPhysX combination and reports them all unusable.
+        config_scan = scan(env_cfg, args)
+        _validate_runtime(config_scan, _get_kit_runtime_sources(config_scan, args))
+        return selected
+    except ImportError:
+        return None
+    except _VALIDATOR_FAILURES as exc:
+        raise ResolutionError(
+            f"validating {task_id!r} could not run: {type(exc).__name__}: {exc}. This is an Isaac Lab API failure,"
+            " not a rejected preset combination."
+        ) from exc
+    except Exception:  # noqa: BLE001 - any other failure means the combination cannot run
+        return None
+    finally:
+        sys.argv = original_argv
+
+
+def resolves(task_id: str, **selection: str | None) -> bool:
+    """Return whether a preset selection can run, for callers that need only that."""
+    return resolve(task_id, **selection) is not None
+
+
+def is_selector(task_id: str, *, physics: str | None = None, renderer: str | None = None) -> bool:
+    """Return whether a preset name is an automatic selector rather than a backend.
+
+    A selector resolves to a config that picks a concrete backend at launch, so it
+    duplicates whichever backend it selects. Recognised by what it resolves to, so a
+    selector added upstream needs no edit here.
+    """
+    resolution = resolve(task_id, physics=physics, renderer=renderer)
+    if resolution is None:
+        return False
+    if physics is not None:
+        return type(resolution.physics).__name__.endswith("AutoCfg")
+    return any(type(cfg).__name__.startswith("_Auto") for cfg in resolution.renderer)
+
+
+def backend_names(specs: list[Any] | None = None) -> frozenset[str]:
+    """Return every name used as a typed physics or renderer preset in the registry.
+
+    Some tasks give an untyped ``PresetCfg`` field a backend name — an observation
+    preset called ``isaacsim_physx``, say — which buckets under ``DOMAIN`` even though
+    it is reported through ``physics=`` / ``renderer=``.
+
+    Resolution cannot identify those: selecting such a preset resolves to the same
+    backend the task already defaults to, so nothing about the resolved config
+    distinguishes it from a genuine domain preset. What does distinguish it is that the
+    name is a typed backend *somewhere* in the registry, which this sweep collects.
+    Deriving it keeps the answer correct as backends are added or renamed, where a
+    hand-written list drifts in both directions.
+
+    Sweeps every registered task, so call once and pass the result around.
+
+    Args:
+        specs: Gym specs to sweep. When ``None``, the whole registry is scanned.
+
+    Returns:
+        Backend names, empty if no task declares a typed preset.
+    """
+    import gymnasium as gym
+
+    from isaaclab_tasks.utils.preset_cli import enumerate_task_presets
+    from isaaclab_tasks.utils.preset_target import PresetTarget
+
+    if specs is None:
+        specs = list(gym.registry.values())
+
+    names: set[str] = set()
+    for spec in specs:
+        # A deprecated task warns on enumeration, and a backend only it declares is on
+        # its way out rather than something callers should recognise.
+        if spec.kwargs.get("deprecated"):
+            continue
+        preset_map = enumerate_task_presets(spec.id)
+        if not preset_map:
+            continue
+        names |= set(preset_map.get(PresetTarget.PHYSICS, ()))
+        names |= set(preset_map.get(PresetTarget.RENDERER, ()))
+    return frozenset(names)
+
+
+def modes(task_id: str, preset_map: dict[Any, list[str]] | None, *, resolve_modes: bool = True) -> tuple[Mode, ...]:
+    """Return the backend combinations for one task.
+
+    A task declaring renderers is expanded across them: reporting a camera task as
+    headless-only omits the thing under test. Domain presets are expanded one at a
+    time, never combined, because presets targeting the same field conflict.
+
+    Args:
+        task_id: Gym task id.
+        preset_map: Output of
+            :func:`~isaaclab_tasks.utils.preset_cli.enumerate_task_presets`, or
+            ``None`` when the config could not be loaded.
+        resolve_modes: When ``True``, each combination is resolved and only usable ones
+            are returned. When ``False``, the full cross product is returned unverified,
+            which is fast but may include combinations that cannot run.
+
+    Returns:
+        Combinations in a deterministic order.
+    """
+    from isaaclab_tasks.utils.preset_target import PresetTarget
+
+    preset_map = preset_map or {}
+    physics_options: tuple[str | None, ...] = tuple(sorted(preset_map.get(PresetTarget.PHYSICS, ()))) or (None,)
+    renderer_options: tuple[str | None, ...] = tuple(sorted(preset_map.get(PresetTarget.RENDERER, ()))) or (None,)
+    domains = tuple(sorted(preset_map.get(PresetTarget.DOMAIN, ())))
+    # ``None`` keeps the task's own default reachable alongside each explicit preset.
+    domain_options: tuple[str | None, ...] = (None, *domains) if domains else (None,)
+
+    found: list[Mode] = []
+    for physics in physics_options:
+        for renderer in renderer_options:
+            for domain in domain_options:
+                if resolve_modes and not resolves(task_id, physics=physics, renderer=renderer, presets=domain):
+                    continue
+                found.append(Mode(physics, renderer, domain))
+    return tuple(found)