Self-Crystalizing Behavior Trees: Amortizing Foundation-Model Reasoning into Deterministic Control for Autonomous Drones
Self-Crystallizing Behavior Trees: Amortizing Foundation-Model Reasoning into Deterministic Control for Autonomous Drones
This research proposes Self-Crystallizing Behavior Trees (SCBT), a method that integrates foundation models into deterministic control systems for autonomous drones. SCBT uses an LLM as a compiler to generate and refine control patches during runtime, which are then crystallized into deterministic behavior trees. The approach demonstrates improved mission success rates compared to unpatched and hand-coded trees, though it faces challenges in correctly attributing certain failure scenarios.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
LLM calls
11
View original evidence
tor. 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
Open source S5LLM calls
26
View original evidence
tor. 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
Open source S5LLM calls
22
View original evidence
tor. 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
Open source S5LLM calls per episode
0.33
View original evidence
tor. 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
Open source S5nodes in crystallized trees
28
View original evidence
tor. 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
Open source S5
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
tor. 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
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org