Integrate Task Engine runtime policies and robot profiles

.github/workflows/main.yml

@@ -564,7 +564,7 @@ jobs:
           dist_info = f"embodichain-{version}.dist-info/"
           required = {
               "embodichain/VERSION",
-              "embodichain/gen_sim/simready_pipeline/configs/gen_config.json",
+              "embodichain/gen_sim/asset_engine/configs/gen_config.json",
               "embodichain_tasks/__init__.py",
               "embodichain_tasks/configs/tasks/manipulation/tableware/pour_water/env.yaml",
               "embodichain_robots/__init__.py",

agent_context/MAP.yaml

@@ -421,25 +421,48 @@ topics:
   status: active
 - id: gen-sim
   title: Gen Sim
-  aliases: [gen sim, Scene Engine, SimReady, scene generation, scene editing, scene-randomization,
-    grouped scene randomization, GPU场景扩增, 批量场景扩增, 组内扩增, 组内场景扩增, intra-group augmentation, 分组场景随机化, 场景生成, 场景编辑, 生成式仿真]
-  keywords: [generate_scene_from_image, edit_scene, SceneExporter, scene-export/v1, SceneRandomizationCfg,
-    randomize_scene, prepare_scene_baseline, scene-randomization, load_layouts, preview_layouts, ingest_one_asset,
-    ParserManager, SessionProcessRegistry, Gradio]
+  aliases: [gen sim, Scene Engine, SimReady, Asset Engine, scene generation, scene editing,
+    articulation generation, articulated asset, task spec, TaskSpec, task template,
+    scene-randomization, grouped scene randomization, GPU场景扩增, 批量场景扩增,
+    组内扩增, 组内场景扩增, intra-group augmentation, 分组场景随机化,
+    资产引擎, 资产引擎关节生成, 任务规范, 任务规格, 关节资产, 场景生成, 场景编辑, 生成式仿真]
+  keywords: [generate_scene_from_image, edit_scene, SceneExporter, scene-export/v1, ingest_one_asset,
+    ParserManager, SessionProcessRegistry, Gradio, articulation, articulated, asset engine,
+    articulation asset generation, TaskTemplate, SceneInstance, ActionWitness,
+    ExpansionManifest, ValidationCertificate, task_spec, semantic_hash, predicate evaluator,
+    task spec cache, TaskSpecCache, ArticulatedGenerationClient, articulated_usdc_utils, SceneRandomizationCfg,
+    randomize_scene, prepare_scene_baseline, scene-randomization, load_layouts, preview_layouts]
   paths: [topics/gen-sim/gen-sim.md]
   source_of_truth:
   - embodichain/gen_sim/scene_engine/cli/start.py
   - embodichain/gen_sim/scene_engine/pipeline/generate.py
   - embodichain/gen_sim/scene_engine/pipeline/edit.py
   - embodichain/gen_sim/scene_engine/pipeline/utils/scene_exporter.py
   - embodichain/gen_sim/scene_engine/pipeline/utils/scene_importer.py
+  - embodichain/gen_sim/asset_engine/pipeline/ingest.py
+  - embodichain/gen_sim/asset_engine/pipeline/articulation.py
+  - embodichain/gen_sim/asset_engine/clients/
+  - embodichain/gen_sim/asset_engine/utils/articulated_usdc_utils.py
   - embodichain/gen_sim/scene_engine/randomization/
-  - embodichain/gen_sim/simready_pipeline/pipeline/ingest.py
   - embodichain/gen_sim/gradio_ui/app_processes.py
+  - embodichain/gen_sim/task_spec/
   - embodichain/gen_sim/gradio_ui/app_env.py
-  watch_paths: [embodichain/gen_sim/, embodichain/cli/main.py, tests/gen_sim/]
+  watch_paths: [embodichain/gen_sim/, embodichain/gen_sim/task_spec/, embodichain/cli/main.py, tests/gen_sim/]
   related_topics: [simulation-system, data-assets, env-framework]
   status: active
+- id: task-spec
+  title: TaskSpec Protocol
+  aliases: [TaskSpec, task spec, task specification, 任务规范, 任务规格]
+  keywords: [TaskSpec, TaskSpecCache, task_spec/v0.1, milestones, semantic_hash,
+    predicate evaluator, task spec cache]
+  paths: [topics/gen-sim/gen-sim.md]
+  source_of_truth:
+  - embodichain/gen_sim/task_spec/
+  - embodichain/gen_sim/task_engine/task_spec.py
+  watch_paths: [embodichain/gen_sim/task_spec/, embodichain/gen_sim/task_engine/task_spec.py,
+    tests/gen_sim/task_spec/, tests/gen_sim/task_engine/test_agent.py]
+  related_topics: [gen-sim, task-programs, motion-planning]
+  status: active
 - id: data-assets
   title: Data Assets
   aliases: [data assets, data path, asset download, 数据资产, 资产下载, 数据路径]

agent_context/topics/atomic-actions/atomic-actions.md

@@ -54,8 +54,31 @@ that commits or invalidates observed effects.
 verified grasp. Its goal optionally frees only the final world yaw; default
 goals remain exact. Heading alternatives are checked as complete motion paths,
 with accepted rows retained independently and selected offsets in diagnostics.
-Held-object transport and Pour keep the binding's configured grasp command; a
-contact-blocked measured aperture does not retain closing effort.
+MoveHeldObject and Pour retain the resolved endpoint's grasp command, including
+invocation overrides, rather than a contact-blocked measured aperture. Place
+retains that command only for held rows, keeps measured posture for free rows,
+and switches to the open command at release. The protected `_held_hand_qpos`
+hook lets a transport specialization own a different hand target without
+changing observations, attachments, failed-row holds or physical effect gates.
+
+Existing-hold `HandOver` resolves its internal exchange object pose and receiving
+grasp together through `_resolve_existing_hold_receive`. The default retains
+the original root-midpoint construction and grasp selection. A specialization
+must return a consistent pair: source transport and the receiving attachment
+are derived from that same pose; the hook never commits effects or steps physics.
+
+Grasp generation retains the original mesh-axis partition. Antipodal's optional
+`pg_grasp.GraspRegionSelector` may select only angle/NMS/collision-qualified fallback
+rows before ranking and top-K, and only when the requested region is empty.
+The default keeps existing custom-service signatures and viable-region RNG
+behavior. Application-specific region interpretation belongs to the provider;
+it cannot recover IK, trajectory, closing or acquisition failure.
+
+Pick projects downstream targets through the protected lazy
+`_iter_downstream_eef_targets` hook before choosing between TCP-roll variants.
+The default preserves object-target ordering, normalization, lift-seeded IK and
+RNG behavior. Specializations may supply candidate-dependent target geometry;
+they do not own IK, roll selection, execution or physical verification.
 
 `SlideGoal.joint_target` optionally supplies a public `SlideJointTarget` with an
 absolute prismatic coordinate and a calibrated axis sign. Slide resolves the

agent_context/topics/atomic-actions/execution.md

@@ -90,6 +90,14 @@ Held-object guards and phase-effect gates are observational:
 Pick gates attachment before lift. Place gates detachment before retract.
 HandOver owns independent source/destination transfer boundaries.
 
+An established held-object relation does not determine the hand's actuator
+target. Shared HandOver preserves the measured source-hand posture for an
+existing hold. Its protected `_existing_hold_hand_qpos` hook lets an application
+specialization retain a bound closing command instead. Contact can keep a
+position-controlled hand short of that target; replacing it with measured angles
+removes closing drive. Target specialization does not replace the shared
+measured attachment, transfer and release verification.
+
 ## Row-local state
 
 Vector environments share a synchronized call and command cursor, but success,

agent_context/topics/gen-sim/acceptance-evidence.md

@@ -0,0 +1,99 @@
+# TaskSpec Acceptance Evidence
+
+## Semantic Boundaries
+
+TaskSpec declares the requested result; recorded execution supplies evidence.
+Neither command completion nor video appearance alone changes the contract.
+Missing evidence is unavailable, not an invented success or a measured failure.
+
+- Events, including action-only pouring, retain their exact historical window.
+  Returning the vessel or moving the receiver afterwards does not erase the event.
+- Stage states are measured at the compiled group's seeded end-step slot.
+- Final states are observed before reset, not copied from an earlier milestone.
+- New E3 instruction interpretation requires explicit `pour_goal_mode`. `motion`
+  verifies alignment, tilt, restoration and release, never liquid transfer.
+  An omitted mode enters bounded interpretation repair. Legacy authored candidates
+  retain their original optional-field semantics; migration creates a new spec.
+
+## Observation Owners
+
+`task_engine/_task_program/` owns these application-level observations:
+
+| Concern | Owner |
+|---|---|
+| Native tilt/restore occurrence | `pour_evidence.py` |
+| Placement/release geometry | `placement_evidence.py`, `task_spec_runtime.py` |
+| Bounded contact observations on Gym's clock | `acceptance_observation.py` |
+| Coordinated acquisition and dual release | `coordinated_placement_evidence.py` |
+| Stable native reference support | `stable_support_evidence.py` |
+| Same-episode comparison and terminal snapshots | `acceptance_audit.py` |
+
+These observers do not step physics, send commands or replace Task Program.
+Runtime profile, assets and execution safety policies remain separate owners.
+
+## Relation Evidence
+
+TaskSpec planar placement uses the union of real native triangle projections.
+A convex hull can invent occupied material across payload concavities. Actual
+reference holes remain excluded; no blanket geometric buffer fills them.
+The execution-side planar gate retains its previous default; exact silhouette
+selection is local to the semantic observer.
+
+A rejected floor projection does not prove violation of a sloping container
+cavity. When released payload geometry overlaps a raised reference and only
+the floor footprint fails, containment remains unavailable until a qualified
+three-dimensional cavity observer exists. Clear misses, penetration and excessive
+support gaps remain failures. Diagnostics retain the native projected areas.
+Current planar containment records carry a per-environment `available` mask;
+matchers recompute it from the measured geometry and release flags. An ambiguous
+row remains unavailable without erasing other rows' final or stage observations.
+
+An E5 composition may collect a native contact timeline through a sensor
+registered in both SimulationManager and Gym's sensor dictionary. The existing
+physics-substep lifecycle supplies samples; all native substeps in each control interval
+are retained with source/topology revisions and explicit capacity/loss reporting.
+Single-ended filtering preserves other counterparts rather than hiding them.
+
+Stable support is distinct from COM projection inside an instantaneous contact
+hull. Its v2 observer requires a complete post-release window and maximum relative
+pose drift within the actual decoded placement preset's bounds. An upward load
+path may traverse known dynamic rigid cargo, but must exist within one native
+substep, not a union of contacts at different times. Each control interval needs
+a measured root path; every substep excludes external positive-impulse contacts
+of the entire coupled cargo component, including frictional side contacts.
+The reference itself may rest on a table. The legacy COM classifier is
+retained for audit. This is sampled native support evidence, not a continuous
+friction-cone or hardware force-closure certificate.
+Zero active contacts cannot distinguish native sleep from absent support and
+remain unavailable unless an independent observation proves a violation.
+Support by a named carrier, whole-footprint containment in a region, and an
+explicit anchor over a region are different contracts, never automatic fallbacks.
+Installing an E5 contact timeline does not replace an E1 `supported_region`'s
+full footprint, contact-gap and exact release-window requirements with support alone.
+
+Coordinated transport observation currently qualifies one environment and an
+explicit configured `on` placement. It binds two distinct resource slots and
+disjoint pad actors, actual two-hand contact in the lift window, an exact
+successful dual release receipt and measured independent open-hand joints.
+Standalone coordinated-hold observation keeps its original contract.
+For an explicit E5 `on` relation, the formal placement evidence is native stable
+support by the named reference. Planar footprint and gap are diagnostics only;
+full coverage cannot certify a released payload hovering above the reference.
+
+## Audit And Qualification
+
+The runner evaluates legacy and current observers before reset in one episode.
+The legacy result is audit-only and cannot override the current selected verdict.
+State guards cover the control clock, step counts, robot qpos, rigid poses and
+articulation poses/qpos. Runtime/program/spec digests bind both reports.
+
+Native meshes and exact terminal poses are saved separately. Existing trajectory
+exports can omit the seeded final slot; never substitute the previous slot.
+Preserve historical specs, videos, receipts and failures. Explicit semantic
+migrations record parent/new hashes and cannot masquerade as same-spec repairs.
+
+Calibration separates human episode-level observations from independently labelled
+per-predicate evidence. Synthetic negatives remain a separate test stratum; they
+do not establish a real-world false-positive rate. Freeze source and input hashes
+before full native regression and report program completion and strict acceptance
+separately.