Language Language English 简体中文

unisim.backend.mjwarp.backend

Host-compatibility implementation of the independent mjwarp backend.

mjwarp is not a MuJoCo backend mode. It uploads a CPU MuJoCo model to mujoco_warp and owns its own device data and host cache. The cache is refreshed exactly at explicit step/reset barriers; host getters only return views into that cache and therefore never trigger an implicit Warp .numpy transfer.

Classes

MjwarpBackend

Independent CUDA backend exposed through the host NumPy profile.

class unisim.backend.mjwarp.backend.MjwarpBackend

Bases: SimBackend

Independent CUDA backend exposed through the host NumPy profile.

State and control cross the host/device boundary only at explicit step/reset barriers with bounded, statically declared transfers. Reset DR writes per-world rows of cold-path-expanded model arrays in place and recomputes derived constants with the graded set_const* family; interval DR stages xfrc_applied pushes/body forces for the next step barrier or kicks root velocity through the reset upload+forward path. A registered pre-step control converter runs on the host before every physics substep (see set_pre_step_control) and may additionally return a per-substep body wrench that composes with the staged interval wrenches; native rendering remains fail-closed. Detached host snapshots support finite MuJoCo-based offline recording.

Parameters:

• scene (SceneCfg)

• num_envs (int)

• sim_dt (float)

• base_name (str | None)

• push_body_name (str | None)

• nconmax (int | None)

• njmax (int | None)

• add_body_sensors (bool)

• unexpected_kwargs (Any)

__init__(scene, num_envs, sim_dt, *, base_name=None, push_body_name=None, nconmax=None, njmax=None, add_body_sensors=False, **unexpected_kwargs)

Parameters:

• scene (SceneCfg)

• num_envs (int)

• sim_dt (float)

• base_name (str | None)

• push_body_name (str | None)

• nconmax (int | None)

• njmax (int | None)

• add_body_sensors (bool)

• unexpected_kwargs (Any)

get_scene_layout()

Return the immutable materialized public addresses, never native handles.

Return type:

CompiledSceneLayout

get_entity_names()

Return materialized physical entity names in the frozen public order.

Return type:

tuple[str, ...]

get_entity_default_state(entity, env_ids=None)

Detached construction/keyframe defaults in selected environment order.

Fields and frames match get_entity_state. Defaults follow immutable variant assignment and never reflect current reset-time randomization. This query does not reset, step or reparse the scene.

Parameters:

• entity (str)

• env_ids (Sequence[int] | ndarray | None)

Return type:

Mapping[str, ndarray]

get_entity_state(entity)

Return detached root_pose/root_velocity/joint_positions/joint_velocities.

Root fields follow EntityStatePatch’s link-origin/world-frame contract. Non-root joint fields use the frozen entity joint order and native generalized widths. Unavailable fields must fail, never return old data as if current. State freshness follows the adapter’s declared profile.

Parameters:

entity (str)

Return type:

Mapping[str, ndarray]

reset_entities(request)

Validate the complete selected-entity request, then commit it once.

Validation failure leaves all state unchanged. A native partial commit failure must fault the backend unless the adapter can actually roll back. Unselected entities/environments and fixed identity remain unchanged.

Parameters:

request (SceneResetRequest)

Return type:

None

cleanup_scene_assets()

Release cold-path scene artifacts owned by the backend.

Return type:

None

close()

Return type:

None

property num_envs: int

Number of vectorized environments.

property model: Any

Return the backend-owned device model, never a MuJoCo backend model.

property num_actuators: int

Number of actuators.

property num_dof_vel: int

Number of joint velocity DoFs, excluding the floating base.

get_actuator_ctrl_range()

Return actuator control ranges.

Return type:

ndarray

Returns:

Array with shape (num_actuators, 2) and columns [low, high].

get_actuator_names()

Return actuator names in control-vector order on the cold path.

Return type:

tuple[str, ...]

get_actuator_joint_names()

Return each actuator’s target single-DoF joint in control-vector order.

Backends must fail closed when an actuator does not target exactly one hinge/slide joint. Manager action terms use this cold-path metadata to map community joint selectors onto the backend control vector without inspecting backend-private model objects.

Return type:

tuple[str, ...]

get_scene_model_file()

Return the materialized scene path for diagnostics, when available.

Return type:

str | None

get_keyframe_qpos(name)

Return the full qpos for a named keyframe, including the floating base.

Parameters:

name (str) – Keyframe name such as "stand" or "home".

Return type:

ndarray

Returns:

Array with shape (nq,).

get_default_qpos()

Return the backend/model default qpos through a stable contract.

Return type:

ndarray

get_default_dof_pos()

Return default joint positions in canonical non-root qpos order.

The returned array is detached, one-dimensional, and excludes floating root coordinates. Backends whose DoF view is actuator-indexed must use that same actuator-target order here.

Return type:

ndarray

get_init_qvel()

Return a zero-initialized qvel vector compatible with set_state_tensor.

Return type:

ndarray

Returns:

Zero-filled qvel array.

get_tensor_execution()

Declare the optional tensor lifecycle without discovering SDKs.

Return type:

TensorExecution

get_tensor_capabilities()

Return fail-closed tensor methods and negotiable state fields.

Return type:

TensorLifecycleCapabilities

bind_packed_reset(*, device=None)

Begin a device-resident transaction with one scatter commit boundary.

Parameters:

device (Any | None)

Return type:

Any

get_public_state_widths()

Return canonical qpos/qvel widths for packed tensor reset layout.

Return type:

PublicStateWidths

get_sensor_names()

Return the complete public named-sensor namespace.

Return type:

tuple[str, ...]

get_sensor_inventory()

Return the complete public named-sensor inventory and widths.

Return type:

tuple[SensorDescriptor, ...]

get_tensor_runtime_diagnostics()

Expose whether fixed-address MJWarp graph replay is active.

Return type:

dict[str, TensorRuntimeDiagnostic]

get_state_views(fields=None, device=None)

Return live Torch views into stable MJWarp device fields.

Parameters:

• fields (tuple[str, ...] | str | None)

• device (Any | None)

Return type:

Mapping[str, Any]

get_sensor_view(name, device=None)

Return a live named-sensor view without exposing sensor offsets.

Parameters:

• name (str)

• device (Any | None)

Return type:

Any

get_tracked_body_views(body_names=None, device=None)

Return four ordered tracked-body blocks from one stable device read.

Parameters:

• body_names (Sequence[str] | None)

• device (Any | None)

Return type:

TrackedBodyStateViews

get_root_state_layout(root_body_name)

Resolve one body’s floating-root columns on the cold path.

Backends must verify that root_body_name owns a free/floating joint; fixed bodies and runtimes without body-to-root metadata fail closed. Name/model lookup is forbidden on reset and step hot paths, so callers cache either the returned layout or the unsupported result during scene materialization.

Parameters:

root_body_name (str)

Return type:

RootStateLayout

get_body_ids(names)

Resolve body/link names to backend integer IDs.

Parameters:

names (Sequence[str]) – Body/link names.

Return type:

ndarray

Returns:

int32 array with shape (len(names),).

Raises:

ValueError – If any name is not found.

get_geom_id(name)

Resolve one geom name through the backend contract.

Parameters:

name (str)

Return type:

int

get_geom_size(name)

Return one geom size vector in backend geom-name order.

get_default_geom_sizes returns the complete table of default sizes, while this selector resolves one geom by name.

Parameters:

name (str)

Return type:

ndarray

get_default_geom_sizes()

Return default geometry sizes, shape (ngeom, 3).

Return type:

ndarray

get_geom_solref()

Return default contact reference parameters, shape (ngeom, 2).

Return type:

ndarray

get_geom_solimp()

Return default contact impedance parameters, shape (ngeom, 5).

Return type:

ndarray

get_dof_damping()

Return default joint damping, shape (nv,).

Return type:

ndarray

get_dof_frictionloss()

Return default joint friction loss, shape (nv,).

Return type:

ndarray

bind_mocap_pose(body_name)

Resolve a mocap body once; unavailable capabilities fail at binding.

Parameters:

body_name (str)

Return type:

MocapPoseBinding

get_body_subtree_ids(root_body_id)

Return body ids in the subtree rooted at root_body_id.

Parameters:

root_body_id (int)

Return type:

ndarray

get_geom_names()

Return backend geom names in backend id order.

Return type:

tuple[str, ...]

get_geom_body_ids()

Return the owning body id for each geom.

Return type:

ndarray

get_geom_contact_masks()

Return per-geom contact type and affinity masks.

Return type:

tuple[ndarray, ndarray]

get_geom_friction()

Return the backend geom-friction table.

Return type:

ndarray

get_gravity()

Return the backend gravity vector.

Return type:

ndarray

get_body_mass()

Return the backend body-mass table.

Return type:

ndarray

get_body_ipos(env_ids=None)

Return body center-of-mass offsets in each body’s local frame, in meters.

body_ipos[b] is the position of body b’s center of mass in that body’s own local coordinate frame; it is neither an inertia tensor nor a whole-subtree COM.

Without env_ids the canonical model default table is returned with shape (nbody, 3) in backend body-id order. It never changes with reset randomization; fixed-variant backends expose their per-environment default baselines through get_reset_term_default() instead.

With env_ids the current effective per-environment values are returned with shape (len(env_ids), nbody, 3), in env_ids order. The values reflect every reset randomization applied so far, including composition with base_com_offset; an environment untouched by a partial reset keeps its previous values. Backends that do not track per-environment inertial offsets fail closed with NotImplementedError for this form.

Parameters:

env_ids (Sequence[int] | ndarray | None)

Return type:

ndarray

get_dof_armature()

Return the backend dof-armature table.

Return type:

ndarray

get_motion_body_ids(names)

Resolve body IDs used by motion datasets.

Motion datasets are generated from MuJoCo, so the returned ids follow the MJCF body order with worldbody as id 0, regardless of the backend-native indexing used by get_body_ids.

Parameters:

names (Sequence[str])

Return type:

ndarray

get_joint_range(*, names=None)

Return joint position limits, excluding the floating base.

Parameters:

names (Sequence[str] | None) – Optional joint names in requested return order. Adapters that support named scalar joints resolve the model mapping internally.

Return type:

ndarray | None

Returns:

Array with shape (num_dof, 2) and columns [low, high], or None when the backend does not expose limits.

When names is provided by a supporting adapter, the result has shape (len(names), 2). Hinge limits are radians, slide limits are meters, and joints without enabled limits return infinities.

get_site_ids(names)

Resolve site names to integer ID arrays.

Parameters:

names (Sequence[str]) – Site names.

Return type:

ndarray

Returns:

int32 ID array with shape (len(names),).

get_joint_dof_indices(names)

Resolve named joint qvel coordinates on the cold metadata path.

Parameters:

names (Sequence[str])

Return type:

ndarray

get_joint_dof_pos_indices(names)

Resolve named single-DoF qpos coordinates excluding the free root.

Parameters:

names (Sequence[str])

Return type:

ndarray

get_joint_dof_vel_indices(names)

Resolve named joint qvel coordinates excluding the free root.

Parameters:

names (Sequence[str])

Return type:

ndarray

get_joint_state_qpos_indices(names)

Resolve named joints to full reset qpos columns.

Parameters:

names (Sequence[str])

Return type:

ndarray

get_joint_state_qvel_indices(names)

Resolve named joints to full reset qvel columns.

Parameters:

names (Sequence[str])

Return type:

ndarray

get_actuator_gains()

Expose immutable model defaults; this does not advertise gain DR support.

Return type:

tuple[ndarray, ndarray]

step_tensor(ctrl, nsteps=1)

Advance MJWarp physics without a control or state host transfer.

Parameters:

• ctrl (Any)

• nsteps (int)

Return type:

dict[str, Any]

set_state_tensor(env_indices, qpos, qvel, randomization=None)

Commit selected generalized state with device-resident scatter kernels.

Parameters:

• env_indices (Any)

• qpos (Any)

• qvel (Any)

• randomization (TensorResetRandomizationPayload | None)

Return type:

dict[str, Any]

set_state_tensor_mask(reset_mask, qpos, qvel, randomization=None, *, full_refresh=False)

Commit full-width tensors selected by a backend-published bool mask.

The public full-width state payload stays fixed-shape. The backend owns mask validation and reset-graph publication, so the caller never probes private row compaction or CUDA graph state.

Parameters:

• reset_mask (Any)

• qpos (Any)

• qvel (Any)

• randomization (TensorResetRandomizationPayload | None)

• full_refresh (bool)

Return type:

dict[str, Any]

get_reset_term_default(term)

Return canonical or fixed-variant authoritative reset defaults.

Parameters:

term (str)

Return type:

ndarray

get_dr_capabilities()

Advertise the per-world model mutation set validated by effect tests.

Return type:

DomainRandomizationCapabilities

apply_interval_randomization(plan)

Apply a scheduled interval randomization plan.

Generic dispatch: each op yielded by plan.iter_ops() is validated against the builtin term specs (custom terms pass through) and routed to the backend-owned handler table returned by _interval_term_handlers(). Op payloads are torch-first: Torch tensors or TensorIntervalPayload values produced by the caller; host adapters convert tensors at their private boundary through unisim.dr._host, and a missing producer-sampled payload fails closed. A term without a handler fails closed with NotImplementedError naming the backend class and the term. Backends that need per-plan prologue/epilogue semantics (for example clearing staged external forces before the ops accumulate) keep a thin override that calls this base implementation.

Parameters:

plan (IntervalRandomizationPlan)

Return type:

None

push_robots(force_range)

Stage caller-owned device push forces for the next tensor step.

The producer owns random generation on the backend Torch device. This boundary receives final full-width forces, validates them on their device, and scatters directly into the live xfrc_applied field; it does not sample on the host, publish the host cache, or mutate the host wrench mirror.

Parameters:

force_range (TensorIntervalPayload)

Return type:

None

apply_body_force(body_ids, force, torque=None)

Accumulate a world-frame wrench on the staged xfrc_applied rows.

The force acts at each target body’s center of mass and the torque is about that center, matching MuJoCo xfrc_applied semantics. The staged wrench applies to every substep of the next step() call and is cleared after it; repeated calls accumulate until consumed.

Parameters:

• body_ids (ndarray)

• force (ndarray)

• torque (ndarray | None)

Return type:

None

materialize()

Finalize cold-path backend resources before reset/step.

Return type:

None

get_play_capabilities()

Return backend-native play/render capabilities.

Return type:

PlayCapabilities

resolve_play_render_plan(*, play_render_mode, play_steps, output_video)

Resolve high-level playback mode into backend-owned render parameters.

Parameters:

• play_render_mode (str | None)

• play_steps (int | None)

• output_video (str | PathLike[str] | None)

Return type:

PlayRenderPlan

run_playback(*, env, initialize, step, num_steps, output_video=None, render_spacing=None, render_offset_mode=None, headless=None, record_video=None, frame_state_getter=None, camera_kwargs=None, debug_overlay_getter=None, on_frame=None)

Execute backend-owned playback for an env wrapper.

camera_kwargs is normalized into CameraCfg at this boundary; unknown mapping keys fail closed with an error naming them.

debug_overlay_getter is an optional per-frame callback returning a sequence with one entry per environment (len == num_envs); each entry is that env’s sequence of DebugPrimitive (None or empty marks an env without overlay) and returning None disables overlays for the frame. Primitive poses are env-local; the renderer applies grid offsets when composing multiple envs. Backends whose get_play_capabilities().supports_debug_overlay is False fail closed with NotImplementedError when this is not None. On the interactive rendering path only backends whose supports_interactive_debug_overlay is True consume it; the others fail closed with NotImplementedError.

on_frame is an optional per-frame video hook called by offline render pipelines before encoding: it receives (frame_index, frame) with the frame an (H, W, 3) uint8 array, and returns a replacement frame of the same shape/dtype or None to keep the original. Backends rendering through a native (non-offline) renderer fail closed with NotImplementedError when this is not None.

Known boundary: env is the owning env wrapper, not a physics-layer concept. Current playback implementations read env-level configuration (e.g. cfg.scene, cfg.ctrl_dt, cfg.render_spacing) and env-owned playback helpers (get_playback_model, get_physics_state_snapshot) that have no backend-native equivalent yet. The parameter stays on this contract until playback asset/config resolution moves onto backend-owned metadata; backends must only use it on the cold playback path.

Parameters:

• env (Any)

• initialize (Any)

• step (Any)

• num_steps (int | None)

• output_video (str | PathLike[str] | None)

• render_spacing (float | None)

• render_offset_mode (str | None)

• headless (bool | None)

• record_video (bool | None)

• frame_state_getter (Any)

• camera_kwargs (CameraCfg | Mapping[str, Any] | None)

• debug_overlay_getter (DebugOverlayGetter | None)

• on_frame (Any)

Return type:

str | None

get_physics_state_layout()

Return the snapshot layout, including the mocap tail when present.

Return type:

PhysicsStateLayout

get_physics_state()

Return a physics snapshot suitable for offline playback/video export.

Rows use the [time, qpos, qvel] layout; backends whose model has mocap bodies append [mocap_pos(nmocap*3), mocap_quat(nmocap*4)] so offline rendering can replay mocap-driven geometry at its recorded pose. Consumers must decode snapshots through get_physics_state_layout().split_state rather than hardcoding column slices.

Return type:

ndarray

get_playback_mocap_state(env_index=0)

Return copied mocap pose arrays for detached visual playback.

Parameters:

env_index (int)

Return type:

tuple[ndarray, ndarray]

get_playback_model(env_index=None)

Return the playback model for a specific env when variants exist.

Parameters:

env_index (int | None) – Optional vectorized environment index.

Return type:

str

Returns:

The backend model object used by playback tooling.