Complete manipulation behavior tree

Summary

Integrate the behavior tree-based pick-and-place pipeline with GraspGen to enable robust, agentic pick-and-place capability in dimOS.

Context

There are two relevant pieces that need to come together:

1. Behavior tree pick-and-place — A py_trees-based orchestration layer exists on the jalaj_BT_pickplace branch (dimos/manipulation/bt/). It has async execution, retry/recovery logic, and grasp verification — but it's out of sync with the current main branch and needs rework.
2. GraspGen — A Docker-based deep-learning grasp generation module already on main (dimos/manipulation/grasping/graspgen_module.py). It's production-ready and supports collision filtering.
3. Heuristic pick-and-place — The current PickAndPlaceModule on main (dimos/manipulation/pick_and_place_module.py) works end-to-end but uses simple heuristic grasps and has no retry/recovery logic.

The goal is to combine the robustness of the BT approach with the grasp quality of GraspGen, wired up to the agentic interface.

Key pointers

Deliverables

• Rebase/rework the BT pick-and-place module onto current main
• Integrate GraspGen as the primary grasp source in the BT pipeline (with heuristic fallback)
• End-to-end pick-and-place working in the xarm_perception_agent blueprint
• Agentic interface — agent can invoke pick, place, pick_and_place skills and receive success/failure feedback
• Basic test coverage for the BT tree structure

Notes

• The BT branch was removed from main due to CI failures — start by understanding what broke before merging code back.
• The existing PickAndPlaceModule skills API (pick, place, drop_on, scan_objects, etc.) is the target interface the agent already knows how to use — preserve it.
• GraspGen runs in Docker and is lazy-initialized on first call. The BT branch already has a GenerateGrasps action node that calls it.

Synced from DIM-747 by stash

Start by comparing dimos/manipulation/pick_and_place_module.py, dimos/manipulation/grasping/graspgen_module.py, dimos/manipulation/blueprints.py, and dimos/manipulation/manipulation_module.py with the dimos/manipulation/bt/ code on the jalaj_BT_pickplace branch. First understand the CI failures that removed the branch, then trace the existing pick, place, and pick_and_place skills in xarm_perception_agent. Done means the BT pipeline works with the existing agentic interface, GraspGen fallback behavior, and basic tree-structure coverage.