Skip to content

Phase 1 results on public trees

Phase 1 has to run on real trees and record the metrics. This page records them for the three example trees and for every public BT.CPP v3 tree in the test fixtures: the BehaviorTree.CPP 3.8.8 examples, the Groot 1 test data and the Nav2 (Humble) navigator trees.

Example trees

Tree Robot Nodes Actions Reduction Transitions Longest guard Approximated (breakage)
DroneMission multirotor 22 10 55% 23 1 ReactiveFallback ×1, ReactiveSequence ×1, RetryUntilSuccessful ×1
NavigateToGoal mobile base 23 11 52% 25 1 none
PickAndPlace arm and gripper 24 13 46% 29 1 RetryUntilSuccessful ×1

Public trees

These numbers are produced by beehave stats and pinned by the snapshot tests in tests/integration/test_public_trees.py, so they cannot drift silently.

Tree Source Nodes Actions Reduction Transitions Longest guard Approximated (breakage)
BehaviorTree (btcpp-subtree-test) btcpp-v3 11 6 45% 13 0 none
ParentIncludeChildIncludeParent (btcpp-nested-includes) btcpp-v3 7 0 100% 1 0 none
MainTree (groot-crossdoor) groot 13 5 62% 13 1 RetryUntilSuccessful ×1
BehaviorTree (groot-show-all) groot 15 1 93% 3 0 ReactiveFallback ×1, ReactiveSequence ×1, Repeat ×1, RetryUntilSuccessful ×1
BehaviorTree (groot-subtree-remapping) groot 7 2 71% 5 0 none
BehaviorTree (groot-key-reordering) groot 18 7 61% 24 3 none
MainTree (nav2-follow-point) nav2 8 3 62% 7 0 GoalUpdater ×1, KeepRunningUntilFailure ×1, PipelineSequence ×1, RateController ×1
MainTree (nav2-replanning-recovery) nav2 18 9 50% 20 1 PipelineSequence ×1, RateController ×1, ReactiveFallback ×1, RecoveryNode ×3, RoundRobin ×1
MainTree (nav2-goal-patience) nav2 24 11 54% 24 1 PathLongerOnApproach ×1, PipelineSequence ×1, RateController ×1, ReactiveFallback ×1, ReactiveSequence ×1, RecoveryNode ×3, RetryUntilSuccessful ×1, RoundRobin ×1
MainTree (nav2-replan-if-invalid) nav2 23 9 61% 22 2 PipelineSequence ×1, RateController ×1, ReactiveFallback ×1, ReactiveSequence ×1, RecoveryNode ×3, RoundRobin ×1
MainTree (nav2-odometry-calibration) nav2 10 8 20% 17 0 Repeat ×1

Nav2 trees are analysed together with Nav2's own nav2_tree_nodes.xml, which declares which of its nodes are conditions:

$ beehave stats navigate_to_pose_w_replanning_and_recovery.xml nav2_tree_nodes.xml

What we learned

  • Coverage: 11 of 11. Every public v3 tree loads, flattens and walks with no manual edits. All four v4 trees are refused with a clear message, including BT.CPP's own Check.xml, which has no BTCPP_format marker and is recognised by its SequenceWithMemory node.
  • Reduction. The graph shows 20% to 100% fewer boxes than the tree has nodes, 61% for the median tree. Trees made mostly of actions in a row, like odometry_calibration, shrink least, which is expected: there is little control logic to hide.
  • Guards stay short. The longest raw guard is 3 terms, before any simplification. The plan's readability target is "about 3 terms after simplification".
  • Breakage is concentrated. Two groups explain every approximated node:
  • Reactive nodes and loops (ReactiveSequence, ReactiveFallback, RetryUntilSuccessful, Repeat, KeepRunningUntilFailure). These are the Phase 2 hard cases already planned.
  • Nav2's custom nodes (RecoveryNode, PipelineSequence, RoundRobin, RateController, GoalUpdater, PathLongerOnApproach). Their semantics are known but not built in. This suggests a Phase 2 feature: let the project config say that a custom node "behaves like" a built-in one (for example PipelineSequence: ReactiveSequence), with a shipped preset for Nav2.
  • Real-world quirks found. BT.CPP's own example subtree_test.xml uses compact subtree references (<CheckStatus/>) that BT.CPP 3.8 rejects at runtime. Beehave reads them as subtrees and warns.

Still open

  • The readability gate: someone who did not write a tree answers "what happens if X fails?" from the graph alone.
  • More real-world trees: the same metrics on larger aerial, mobile and manipulation trees contributed by users.