feat(atomic-actions): extend handover coordination and slide targets

[yuecideng]

Description

This PR isolates the Atomic Actions and direct grasp-generation changes from the larger action-engine branch. It extends held-object continuity, handover/release behavior, coordinated dual-arm release, articulated slide targets, scene-relative pose goals, and contact-aware planning diagnostics.

The change is based directly on main. Task Engine orchestration, Task Program bundle generation, TaskSpec, Asset Engine, and Scene Engine changes remain outside this PR.

Key behavior changes:

• SceneEntityPose supports world-frame displacement and orientation overrides.
• Pick, AxisAlign, and Place stop treating expected contact motion as an external scene revision after their configured commit boundary.
• MoveHeldObject supports ordered object-pose waypoints and bounded final world-yaw alternatives while preserving the verified grasp.
• HandOver supports transfer-only execution, existing-held continuation, optional release-at-target behavior, and explicit source retreat phases.
• CoordinatedPickment supports synchronized release and post-release retreat.
• SlideGoal supports absolute prismatic joint targets with calibrated axis signs and same-timestamp observation checks.
• GraspKit collision and antipodal filtering is tightened for the expanded action routes.
• Atomic Actions documentation, API export coverage, and focused tests are updated.

Dependencies: none; this PR targets main.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• Enhancement (non-breaking change which improves an existing functionality)
• New feature (non-breaking change which adds functionality)
• Breaking change (existing functionality will not work without user modification)
• Documentation update

Validation

• python -m compileall -q embodichain/lab/sim/atomic_actions embodichain/toolkits/graspkit
• Black check passed for changed Python files.
• python docs/scripts/check_api_docs.py passed: 2415/2415 exports documented.
• git diff --check passed.
• pytest -q tests/sim/atomic_actions tests/toolkits was attempted, but collection is blocked in the current environment by the installed DexSim package: dexsim.types does not export RTRenderMode, which is imported by main's rendering configuration before the tests load.

[greptile-apps]

Confidence Score: 4/5

[Medium risk] Extends atomic action primitives with new motion capabilities.

The PR does not yet appear safe to merge because an existing-held HandOver can continue toward a stale exchange height when its scene-referenced target moves.

Findings

1.  Exchange target stops being monitored ▶

Fix with agent prompt

### Issue 1
embodichain/lab/sim/atomic_actions/primitives/hand_over.py:1022-1026
When the existing-held route uses a scene entity as `goal.target_pose`, its pose determines the exchange height. This cutoff disables monitoring from the start. If that entity moves during transfer, execution continues toward the stale height instead of replanning.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

The PR expands atomic-action handover, coordinated release, slide targeting, held-object transport, and grasp planning. Since the previous review, it also narrows AxisAlign’s scene-monitoring cutoff, restores full Place destination monitoring, and makes MoveJoints agent-visible.

• The Place monitoring concern is fixed, and distinct AxisAlign grasp targets remain monitored.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Scene and robot observations] --> B[Plan atomic action]
  B --> C{Action route}
  C --> D[Handover and release]
  C --> E[Slide joint target]
  C --> F[Held-object waypoints]
  D --> G[Execute timed commands]
  E --> G
  F --> G
  G --> H[Verify effects and update task state]

Loading

_{Reviews (3) · Last reviewed commit: "fix(atomic-actions): keep dynamic target..."}

[greptile-apps]

Exchange target stops being monitored When the existing-held route uses a scene entity as goal.target_pose, its pose determines the exchange height. This cutoff disables monitoring from the start. If that entity moves during transfer, execution continues toward the stale height instead of replanning.

Prompt To Fix With AI

This is a comment left during a code review.
Path: embodichain/lab/sim/atomic_actions/primitives/hand_over.py
Line: 1016-1018

Comment:
**Exchange target stops being monitored** When the existing-held route uses a scene entity as `goal.target_pose`, its pose determines the exchange height. This cutoff disables monitoring from the start. If that entity moves during transfer, execution continues toward the stale height instead of replanning.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.