refactor(task-engine): execute GenSim semantic graphs via Task Program

docs/source/api_reference/embodichain/embodichain.gen_sim.task_engine.rst

@@ -0,0 +1,132 @@
+embodichain.gen_sim.task_engine
+================================
+
+The Task Engine turns instructions and generated-scene evidence into immutable
+semantic task graphs.  Its execution workflow delegates Semantic Calls to the
+canonical Task Program runtime; it does not ground atomic goals or issue robot
+commands itself.
+
+Public facade
+-------------
+
+.. automodule:: embodichain.gen_sim.task_engine
+   :members:
+   :imported-members:
+   :no-index:
+
+Interpretation and planning contracts
+-------------------------------------
+
+.. automodule:: embodichain.gen_sim.task_engine.agent
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.cli
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.config
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.contracts
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.interpretation
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.ontology
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.state_machine
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.workflow_contracts
+   :members:
+   :no-index:
+
+Scene orchestration
+-------------------
+
+These modules preserve canonical scene identity while separating generated
+authoring data, conservative planning evidence, and live simulator bindings.
+
+.. automodule:: embodichain.gen_sim.task_engine.orchestration
+   :members:
+   :imported-members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.orchestration.artifacts
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.orchestration.contracts
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.orchestration.coordinator
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.orchestration.legacy_scene
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.orchestration.scene_adapter
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.orchestration.scene_source
+   :members:
+   :no-index:
+
+Scene analysis and inspection
+-----------------------------
+
+.. automodule:: embodichain.gen_sim.task_engine.scene
+   :members:
+   :imported-members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.scene.conservative_graph
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.scene.contracts
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.scene.feasibility
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.scene.final_inspection
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.scene.scene_engine_v1
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.scene_backend
+   :members:
+   :no-index:
+
+Execution workflow
+------------------
+
+The workflow owns run directories, bounded orchestration attempts, and
+tensor-free reports.  Physical execution remains behind the configured Task
+Program subprocess boundary.
+
+.. automodule:: embodichain.gen_sim.task_engine.run_directory
+   :members:
+   :no-index:
+
+.. automodule:: embodichain.gen_sim.task_engine.workflow
+   :members:
+   :no-index:

docs/source/api_reference/embodichain/embodichain.lab.task_program.compiler.rst

@@ -13,6 +13,7 @@ embodichain.lab.task_program.compiler
 
       CompiledArticulationJointPositionValidator
       CompiledBarrier
+      CompiledObjectNearRelativeTargetValidator
       CompiledObjectNearTargetValidator
       CompiledParallelBlock
       CompiledParallelBranch
@@ -26,4 +27,3 @@ embodichain.lab.task_program.compiler
       CompiledTaskProgramValidator
       TaskProgramCompileError
       TaskProgramCompiler
-

docs/source/api_reference/embodichain/embodichain.lab.task_program.rst

@@ -24,6 +24,7 @@ embodichain.lab.task_program
       BarrierCfg
       WaitStablePostCfg
       ObjectNearTargetValidatorCfg
+      ObjectNearRelativeTargetValidatorCfg
       ArticulationJointPositionValidatorCfg
       TaskProgramCompiler
       CompiledTaskProgram
@@ -98,6 +99,9 @@ Schema
 .. autoclass:: ObjectNearTargetValidatorCfg
    :members:
 
+.. autoclass:: ObjectNearRelativeTargetValidatorCfg
+   :members:
+
 .. autoclass:: ArticulationJointPositionValidatorCfg
    :members:
 

docs/source/api_reference/index.rst

@@ -79,4 +79,5 @@ documentation. CI runs this same checker after style checks and before tests.
 .. toctree::
    :maxdepth: 1
 
+   embodichain/embodichain.gen_sim.task_engine
    public_api

docs/source/api_reference/public_api.rst

@@ -228,6 +228,125 @@ embodichain.gen_sim.gradio_ui.gradio_app
 
    main
 
+embodichain.gen_sim.task_engine.orchestration.grounding
+--------------------------------------------------------
+
+Task-conditioned grounding selects canonical scene UIDs from a redacted
+semantic inventory. It does not construct physical goals or action
+invocations.
+
+.. currentmodule:: embodichain.gen_sim.task_engine.orchestration.grounding
+
+.. autosummary::
+
+   GroundingCaller
+   GroundingResult
+   ground_scene_references
+
+embodichain.gen_sim.task_engine.orchestration.scene_assets
+-----------------------------------------------------------
+
+Generated GLB assets are normalized into renderer-safe runtime geometry while
+the source files and their provenance remain unchanged.
+
+.. currentmodule:: embodichain.gen_sim.task_engine.orchestration.scene_assets
+
+.. autosummary::
+
+   normalize_scene_assets
+
+embodichain.gen_sim.task_engine.orchestration.scene_inventory
+--------------------------------------------------------------
+
+The structural inventory preserves source semantics and validates only
+explicit task/scene compatibility. It deliberately avoids language matching
+and physical goal grounding.
+
+.. currentmodule:: embodichain.gen_sim.task_engine.orchestration.scene_inventory
+
+.. autosummary::
+
+   SceneEntity
+   SceneInventory
+   validate_source_compatibility
+   validate_target_compatibility
+
+embodichain.gen_sim.task_engine.orchestration.source_scene
+-----------------------------------------------------------
+
+Source-scene utilities resolve supported exports and normalize their stable
+identities, paths, transforms, and conservative physics metadata for Task
+Engine planning and simulator loading.
+
+.. currentmodule:: embodichain.gen_sim.task_engine.orchestration.source_scene
+
+.. autosummary::
+
+   PreparedScene
+   ResolvedSceneSource
+   is_prompt2scene_export
+   prepare_scene
+   resolve_gym_config_path
+   resolve_source_scene
+
+embodichain.gen_sim.task_engine.reporting
+------------------------------------------
+
+Tensor-free Task Engine reports retain the canonical Task Program runtime
+result and its per-environment semantic outcomes without reconstructing
+physical execution state.
+
+.. currentmodule:: embodichain.gen_sim.task_engine.reporting
+
+.. autosummary::
+
+   EXECUTION_REPORT_FILENAME
+   TASK_PROGRAM_EXECUTION_REPORT_SCHEMA
+   TaskProgramExecutionReport
+   validate_execution_report
+   write_execution_report
+
+embodichain.gen_sim.task_engine.semantic_graph
+-----------------------------------------------
+
+Semantic task graphs are immutable, JSON-safe planning artifacts whose nodes
+contain canonical Semantic Calls and exclude grounded action data.
+
+.. currentmodule:: embodichain.gen_sim.task_engine.semantic_graph
+
+.. autosummary::
+
+   SEMANTIC_TASK_GRAPH_SCHEMA
+   SemanticTaskGraph
+   semantic_task_graph_hash
+   validate_semantic_task_graph
+
+embodichain.gen_sim.task_engine.semantic_planner
+-------------------------------------------------
+
+The semantic planner lowers validated Task Engine candidates and canonical
+scene bindings into provider-free semantic task graphs.
+
+.. currentmodule:: embodichain.gen_sim.task_engine.semantic_planner
+
+.. autosummary::
+
+   SemanticTaskPlanner
+   UnsupportedSemanticCapabilityError
+
+embodichain.gen_sim.task_engine.task_program_bundle
+----------------------------------------------------
+
+Bundle generation materializes and preflights a fingerprint-bound configured
+Task Program deployment from a semantic task graph and prepared scene.
+
+.. currentmodule:: embodichain.gen_sim.task_engine.task_program_bundle
+
+.. autosummary::
+
+   TaskProgramBundlePaths
+   generate_task_program_bundle
+
 embodichain.gen_sim.scene_engine.core.scene_edit_plan
 -----------------------------------------------------
 
@@ -619,6 +738,7 @@ embodichain.lab.sim.atomic_actions.control
    GRASP_COMMAND
    JointPositionCommand
    OPEN_COMMAND
+   PARK_COMMAND
 
 embodichain.lab.sim.atomic_actions.core
 ---------------------------------------
@@ -1296,6 +1416,16 @@ embodichain.lab.sim.solvers.null_space_posture_task
 
    NullSpacePostureTask
 
+embodichain.lab.sim.solvers.pytorch_solver
+------------------------------------------
+
+.. currentmodule:: embodichain.lab.sim.solvers.pytorch_solver
+
+.. autosummary::
+
+   PytorchSolverCfg
+   PytorchSolver
+
 embodichain.lab.sim.solvers.pink_solver
 ---------------------------------------
 

docs/source/overview/sim/atomic_actions/builtin_actions.md

@@ -231,9 +231,10 @@ control-part adapter resolves current joint-backed endpoints through
 `Robot.control_parts`; custom adapters may instead return mobile, whole-body, or
 other runtime targets.
 
-`MoveJoints` is intentionally `agent_visible=False`: it is useful for home,
-recovery, calibration, and scripted postures, but is not exposed to an Action
-Agent by default.
+`MoveJoints` is also the canonical implementation for home, recovery,
+calibration, and other embodiment-named postures. Semantic callers should
+constrain those uses through a registered call and keep the named target in the
+robot profile.
 
 ## Shared goal and configuration rules
 
@@ -253,7 +254,11 @@ Explicit pose tensors use `(4, 4)` or `(B, 4, 4)`. Waypoint-capable fields in
 `EndEffectorPoseGoal` and `PlaceGoal` also accept `(B, N, 4, 4)`.
 `SceneEntityPose` resolves to the latest `(B, 4, 4)` pose from each
 `SceneSnapshot`, checks optional perception confidence, and registers that
-entity as a recovery dependency.
+entity as a recovery dependency. `world_displacement` keeps a translation in
+the world frame after local composition. `world_orientation` replaces the
+tracked entity's rotation before applying `relative_pose`, allowing a target to
+track a moving reference position while retaining a grounded world-frame
+orientation.
 
 | Skill / field | `SceneEntityPose` accepted | Automatic scene-motion replan |
 |---|---:|---:|
@@ -379,7 +384,7 @@ than an EEF pose.
 | Motion | joint planning/interpolation from observed qpos; supports joint waypoints |
 | Completion | `JOINT_GOAL_REACHED` |
 | Effect | none |
-| Agent visibility | hidden by default (`agent_visible=False`) |
+| Agent visibility | visible |
 
 `target` accepts an explicit qpos tensor with shape `(control_dof,)`,
 `(B, control_dof)`, or `(B, N, control_dof)`, or a non-empty string resolved