feat(tasks): add declarative expert program vertical slices

[yuecideng]

Description

Stack

• Layer: 15/28
• Base: feat/mllm-expert-program-frontend
• Depends on: feat(agents): add strict expert program frontend #498
• Followed by: feat(tasks): add declarative task slices and rollout validation #500

Migrate repeated Cube Pick/Place and Open Drawer to declarative Expert Programs with typed scene/profile integration. Task classes no longer own waypoint math, invocation assembly, or demo-generation overrides.

The two vertical slices prove configuration-only task expansion when the shared semantic capability already exists, while keeping task validation and physical settling explicit.

Refs #471
Refs #474

Type of change

• New feature (non-breaking change which adds functionality)

Screenshots

Not applicable.

Validation

• Focused coverage: tests/gym/envs/expert_program/test_task_vertical_slices.py, cube/open-drawer task tests, and package-data tests
• Final affected-suite regression on the stack tip: 1215 passed, 2 skipped, 8 deselected
• Changed Python files pass Black 26.3.1; the Sphinx build and rollout-report drift check pass at the stack tip

Checklist

• Changed Python files pass Black 26.3.1.
• Corresponding public/design documentation is included in this stack.
• Tests cover the affected behavior.
• No dependency update is required.

[greptile-apps]

Greptile Summary

The PR migrates repeated cube pick/place and Open Drawer demonstrations from task-owned motion generation to declarative Expert Programs backed by typed scene and robot-profile bindings.

• Adds packaged JSON/YAML Expert Programs and selects them from the corresponding Gym configurations.
• Replaces task-specific trajectory assembly with shared simulation Expert Program adapters.
• Adds reset settling for the cube task and declarative post-settling and validation behavior.
• Expands focused task, vertical-slice, and package-data coverage.
• Updates architectural and user documentation for the semantic runtime, effect monitoring, articulation operations, and Expert Programs.

Confidence Score: 5/5

The PR appears safe to merge, with no concrete blocking or independently actionable non-blocking issue identified.

The declarative program identifiers align with their typed task bindings, configuration paths resolve through the supported loading paths, and the investigated initialization, binding, and test-security concerns did not establish observable failures.

Important Files Changed

Filename	Overview
embodichain_tasks/embodichain_tasks/multi_segments/cube_pick_place.py	Replaces task-owned pick/place trajectory and validation logic with typed cube scene/profile declarations and the shared Expert Program adapter.
embodichain_tasks/embodichain_tasks/tableware/open_drawer.py	Replaces the fixed drawer-opening trajectory with typed articulation geometry, robot resources, command presets, and shared semantic runtime integration.
embodichain_tasks/configs/expert_program/multi_segments/repeated_cube_pick_place.yaml	Declares three repeated cube pick/place segments with cyclic targets, settling, and target-proximity validation.
embodichain_tasks/configs/expert_program/tableware/open_drawer.json	Declares one articulation-operation segment targeting the drawer’s named open state.
embodichain_tasks/configs/gym/multi_segments/cube_pick_place.json	Selects the packaged cube Expert Program and adds a bounded reset-settling event.
tests/gym/envs/expert_program/test_task_vertical_slices.py	Adds focused coverage for the two declarative task integrations and their shared runtime path.
tests/test_expert_program_package_data.py	Verifies that the added Expert Programs are included and decode correctly from staged package data.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    A[Gym task config] --> B[Packaged Expert Program]
    C[Task scene binding] --> D[Simulation Expert Program adapter]
    E[Task robot profile binding] --> D
    B --> D
    D --> F[Semantic compiler and runtime]
    F --> G[Buffered commands through env.step]
    G --> H[Effect monitoring]
    H --> I[Settling and validation metadata]

Loading

_{Reviews (1): Last reviewed commit: "feat(tasks): add declarative expert prog..." | Re-trigger Greptile}

[Copilot]

Pull request overview

This PR migrates two existing simulation tasks (multi-segment cube pick/place and open drawer) from task-owned waypoint/motion planning code to declarative Expert Programs (JSON/YAML) backed by typed scene + robot-profile bindings, validating that tasks can be expanded “configuration-only” when the underlying semantic capability already exists.

Changes:

• Convert MultiSegmentsCubePickPlaceEnv and OpenDrawerEnv to use ExpertProgramEnvironmentMixin + create_simulation_expert_program_adapter, leaving tasks responsible only for scene/profile declarations.
• Add packaged Expert Program resources (repeated_cube_pick_place.yaml, open_drawer.json) and update Gym configs to reference them via expert_program_path.
• Add focused vertical-slice tests and update documentation/context to cover Expert Programs, effect evidence/monitoring, and articulation operation semantics.

Reviewed changes

Copilot reviewed 18 out of 18 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/test_expert_program_package_data.py	Verifies setuptools stages Expert Program JSON/YAML and that staged resources decode via installed config paths.
tests/gym/envs/tasks/test_open_drawer.py	Validates OpenDrawer task registration, config wiring to packaged program, adapter construction, and (optional) real-sim execution evidence.
tests/gym/envs/tasks/test_multi_segments_cube_pick_place.py	Validates cube task registration, config wiring to packaged program, adapter construction, and profile/settle configuration.
tests/gym/envs/expert_program/test_task_vertical_slices.py	End-to-end non-physical vertical slices: strict decode/compile, lazy lifecycles, settle/validator metadata, and task “no motion override” guarantees.
embodichain_tasks/embodichain_tasks/tableware/open_drawer.py	Replaces task-local motion generation with typed scene/profile declarations for operate_articulation.
embodichain_tasks/embodichain_tasks/multi_segments/cube_pick_place.py	Replaces task-local lazy planning/settling with typed scene/profile declarations plus packaged repeated-cube Expert Program.
embodichain_tasks/configs/gym/open_drawer/cobot_magic_3cam.json	Adds expert_program_path to select the packaged drawer program.
embodichain_tasks/configs/gym/multi_segments/cube_pick_place.json	Adds expert_program_path, moves settling to reset events, and trims extensions to semantic-only knobs.
embodichain_tasks/configs/expert_program/tableware/open_drawer.json	Adds the declarative Open Drawer Expert Program (segment invoking operate_articulation).
embodichain_tasks/configs/expert_program/multi_segments/repeated_cube_pick_place.yaml	Adds the declarative repeated cube pick/place Expert Program (repeat of segment with post-policy + validator).
docs/source/tutorial/atomic_actions.rst	Documents scene-dependency monitoring cutoffs via scene_dependency_monitor_until.
docs/source/overview/sim/index.rst	Adds navigation guidance for when to use Expert Programs.
docs/source/overview/sim/atomic_actions/robot_skill_profiles.md	Extends docs for effect monitors, endpoint metadata, and Expert Program integration patterns.
docs/source/overview/sim/atomic_actions/index.md	Adds Expert Programs doc to the atomic-actions overview and clarifies effect evidence vs decision responsibilities.
docs/source/overview/sim/atomic_actions/builtin_actions.md	Documents operate_articulation and its dependency-monitor cutoff semantics.
docs/design/declarative_expert_program_plan.md	Updates project design plan status/progress notes for the stacked implementation.
agent_context/topics/atomic-actions/atomic-actions.md	Updates agent-facing context with semantic runtime + Expert Program details.
agent_context/MAP.yaml	Expands topic keywords and file pointers for Expert Program / semantic runtime content.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[yuecideng]

Folded into #500 during stacked-PR consolidation. Its commits remain included in #500; the remote branch is retained for traceability.