Self-crystallizing behavior trees improve drone autonomy and reliability
Self-Crystallizing Behavior Trees: Amortizing Foundation-Model Reasoning into Deterministic Control for Autonomous Drones
Robotics
Summary
Drones often rely on fixed control programs that can fail when unexpected problems arise, like sudden wind or battery issues. The authors developed a system called Self-Crystallizing Behavior Trees (SCBT) that lets a large AI model suggest fixes only when needed, then turns those fixes into permanent, reliable control steps. This reduces the slow AI calls during flight and results in safe, predictable drone behavior that can adapt to new challenges. Their system reached much higher success rates in simulated drone missions compared to unpatched control programs.
What this means in practice
- •For drone system engineers: Integrate AI-guided patching that crystallizes into deterministic control trees to improve drone autonomy under unexpected conditions.
- •For autonomous vehicle control teams: Use foundation models as compilers for runtime fixes to reduce non-determinism and latency in safety-critical vehicle control software.
Authors
Behnam Moradi
Abstract
Classical robot control, Finite State Machines and Behavior Trees (BTs), is deterministic and inspectable but fixed at design time: it fails on situations its designers did not anticipate. Foundation models can synthesize new control logic at runtime, at a cost in latency and non-determinism hard to accept in safety-critical aerial systems. Existing LLM-driven approaches keep the model in the loop permanently and discard what it produces. We propose Self-Crystallizing Behavior Trees (SCBT), in which the foundation model is a compiler rather than a runtime. When a BT node fails, an LLM proposes a minimal, mission-agnostic subtree patch; CMA-ES tunes its parameters in simulation; a contextual bandit decides whether a stored patch or a new LLM call is worth its cost; and a three-stage admission pipeline (static check, invariant-monitored replays, an empirical margin over the unpatched tree) admits only patches that demonstrably recover. Admitted patches are stored in a fleet library and, once the LLM call rate decays, crystallized into a deterministic BT that no human wrote. We implement SCBT on Microsoft Project AirSim with a schema-constrained LLM pipeline (Gemini 3.1 Pro) and a lockstep simulator. Over three seeds of a 20-episode learning phase with wind-gust and battery failures, SCBT used $11\pm1.4$ LLM calls against 26 for whole-tree regeneration and 22 without a library, with 0.33 calls per episode in the second half. The crystallized trees, 28 and 33 nodes with two patches each, reached $97\pm2\%$ mission success with no learning running, against 12\% for the unpatched tree and 98\% for a hand-coded expert tree, and reproduce decision traces exactly when re-ticked offline. Two negative results: replay admission rejected useful wind recoveries by attributing wind-induced geofence excursions to the patch, and a static-only ablation generalized better to held-out failures.