Add a behavior-tree recovery example (manipulation/37)

[rsasaki0109]

What

A new example, examples/manipulation/37_behavior_tree_recovery.py, that drives the tabletop pick with a reactive behavior tree instead of imperative if/else recovery.

The repo's whole thesis is what the robot does after it fails. Behavior trees are the pattern roboticists actually reach for to structure that recovery — BehaviorTree.CPP, py_trees, the ROS 2 Nav2 BT Navigator — yet the repo had no BT example. This fills that gap with the smallest honest version.

The lesson

Same task, same world (Tabletop2D) as the hero loop 01_pick_and_retry.py, but the recovery is declarative:

Fallback "pick the block"
├── Condition  object_in_gripper?            # already holding -> whole tree SUCCESS
└── Fallback   "grasp or recover"
    ├── Sequence "confident grasp"
    │   ├── Condition  belief_confident?
    │   └── Action     grasp_at_belief
    └── Action     relook_to_refine           # recovery: scan a new viewpoint

A grasp miss is not handled by a special branch — it just grows the belief radius, flips belief_confident? to false, and the same fallback re-runs active perception before retrying. Occlusion and grasp-miss recovery share one declarative branch. Recovery is a property of the tree's shape, not a hand-written routine.

Contents

• Minimal three-status BT engine (Status / Sequence / Fallback / Condition / Action) inline — no external library; one tick yields exactly one env action (the RUNNING leaf).
• A belief-tracking BehaviorTreeAgent (mean + shrinking radius, EMA over detections), mirroring 01's belief so the contrast is purely structural.
• Two smoke tests: root reports SUCCESS on the hero seed; a miss-heavy seed falls back through relook_to_refine after a grasp_miss and still succeeds.
• examples/README.md row + a full examples/manipulation/README.md section; example count 39→40 and test count 111→113 in README.md / docs/status.md.

Verification

• Headless across seeds 0–7: all succeed; re-looks interleave with grasps, retries on miss (e.g. seed 4 → 3 grasps / 2 retries, seed 3 → clean).
• Full suite green (126 passed); matplotlib render path confirmed under Agg.

References

• M. Colledanchise & P. Ögren, Behavior Trees in Robotics and AI: An Introduction, CRC Press 2018 (arXiv:1709.00084).
• BehaviorTree.CPP, py_trees.

🤖 Generated with Claude Code