cs.ROSep 4, 2026

Adaptation Needs in Robotic Systems: Assessing Behavior Trees and Their Enhancement

Authors: Mehran Rostamnia, Gianluca Filippone, Ricardo Caldas, Patrizio Pelliccione

Organizations: Gran Sasso Science Institute (GSSI), L’Aquila, Italy

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Robotic systems increasingly operate in dynamic, uncertain, and open-ended environments, where design-time assumptions may no longer hold, and adaptation becomes necessary to maintain effective and safe operation. Behavior Trees (BTs) are widely used in robotic control architectures due to their modularity, readability, and reactivity. This raises a central question: are BTs sufficient to meet the adaptation needs of modern robotic systems? This paper investigates this question through a literature-driven study complemented by empirical validation. First, we derive a classification of robotic adaptation needs from the literature, organizing them into six categories: Knowledge, Perception, Actuation, System, Mission, and Environment. Then, we analyze the capabilities and limitations of classical BTs with respect to these needs. Then, we characterize BT-based approaches for adaptation from the existing literature and organize them into four primary families, i.e., generation, extension, evolution, and refinement, including approaches that combine multiple families. Our analysis shows that the modularity, flexibility, and reactivity of classical BTs are insufficient for adaptation needs involving runtime restructuring, reasoning under uncertainty, mission reinterpretation, learning, or integration with external knowledge and planning mechanisms. Enhanced BT approaches address several of these limitations, but to different extents and often with limitations of their own. Our findings relate adaptation needs to both the capabilities and limitations of classical and enhanced BTs, providing guidance on when classical BTs are sufficient, when enhanced mechanisms are needed, and which challenges remain or emerge for adaptive robotic control architectures.

Figures & tables

Fig. 1: Research design for identifying adaptation needs, assessing classical BTs, characterizing BT-based adaptation approaches, and deriving research gaps.
Fig. 2: Participants’ background (46 respondents). (a) Self-reported experience in robotics and BTs. (b) Robotics domains in which respondents have worked. Percentages in (b) do not sum to 100% because respondents could select multiple domains.
Abstraction• Definition: Knowledge abstraction refers to the omission or oversimplification of real-world details of the robotic system itself, its operational environment, or the robot-environment interaction, in the knowledge representation maintained by the system [ 30 , 24 , 8 ] . Knowledge abstraction is one of the reasons for the reality gap [ 31 ] . • Trigger : Triggered by a discovered mismatch between the abstracted knowledge and the actual system state. • Object(s) of Adaptation: Robot’s knowledge representation to reduce the reality gap [ 31 ] , the unpredictability of the robot action’s outcome, or in its assessment of the current system environment state. • Example: To simplify the robot-world interaction knowledge, details of the wheel-ground contact may be omitted, introducing model mismatch that can lead to accumulated localization (pose) drift during mission execution [ 32 ] .
Incompleteness• Definition: Knowledge incompleteness refers to the lack or unavailability of some required piece of information about the robotic system itself, its operational environment, or the actions/tasks to be performed [ 33 ] . Knowledge incompleteness is one of the reasons for the reality gap [ 31 ] . • Trigger : Triggered by direct observation (e.g., through sensors) or inference mechanisms (e.g., analysis of its own behavior, benchmarking against expectations). • Object(s) of Adaptation: The robot’s knowledge to reduce unpredictable system behavior, which may be consequent to strange behaviors, mission failure, or system error. • Example: A rover may encounter a new type of terrain that was not included in its original real-world environmental knowledge [ 20 ] . Since the terrain characteristic is unknown, the rover cannot accurately predict the effects of its actions (e.g., slippage).
Inaccuracy• Definition: Knowledge inaccuracy refers to a difference between the knowledge maintained by the robot and the actual state of the robotic system itself, its operational environment, or the robot-environment interaction, stemming from an inaccurate implementation of knowledge [ 34 ] . Knowledge inaccuracy is one of the reasons for the reality gap [ 31 ] . • Trigger : Triggered by inconsistencies between the predicted outcome based on the robot’s knowledge and the actual observed outcomes in the real world. • Object(s) of Adaptation: The robot’s internal knowledge about system properties, environment characteristics, or task-related parameters used for perception, planning, and control. • Example: A robot may grasp an object using an inaccurate estimation of object mass or friction properties [ 35 ] . Since the internal knowledge does not represent the real object accurately, it may slip or fall from the robot’s gripper.
Drift• Definition: Knowledge drift refers to the mismatch between the knowledge maintained by the robot and the actual state of the robotic system or the physical world in which it operates [ 8 ] . It occurs when the initial assumptions about the knowledge used by the robot become outdated due to changes in the system, environment, or operational context over time. Knowledge drift is one of the reasons for the reality gap [ 31 ] . • Trigger : Triggered by direct observation (e.g., through sensors) or inference mechanisms (e.g., analysis of own behavior, benchmark with expectations). • Object(s) of Adaptation : Knowledge to reduce the reality gap [ 31 ] , different strategies, controllers, etc., to compensate for potential behavior degradations. • Example: Knowledge drifts may manifest when a robot’s internal map indicates a corridor is free while its sensors detect an unexpected obstacle that had been added by the robot itself or other environmental factors during the mission execution, revealing a mismatch between the robot’s model and real-world phenomena and leading to degraded or unreliable behavior [ 36 ] .
TABLE I: Knowledge-driven adaptation needs.
Multiplicity• Definition: Perception multiplicity refers to heterogeneity and possible inconsistency of information coming from multiple sources of information (e.g., cameras, sensors, LiDAR), which may represent the same real-world entities or events in different ways or provide conflicting observations [ 38 ] . • Trigger: Triggered by the detection of inconsistent or conflicting information about the same entity, event, or state when combining observations from multiple sensors or information sources. • Object(s) of Adaptation: The robot’s data fusion mechanisms and internal knowledge used to integrate and reconcile information from multiple sensing sources. • Example: A robot may receive different estimations of the pedestrians’ positions from its camera and LiDAR sensors [ 39 ] . Since the two perception sources perceive and represent the environment differently, their observations may conflict.
Inaccuracy• Definition: Perception inaccuracy refers to the discovery of incomplete, noisy, or unreliable observations of the environment, which stems from the inherent imperfection of sensors. This inaccuracy can stem either from internal stimuli (the system itself) or external stimuli (the system detects a problem that does not exist, or fails to detect a problem that actually occurs) [ 40 ] . • Trigger: Triggered by the detection of unreliable, noisy, or incomplete sensor observations, which may originate from internal sensing issues or external conditions that affect the sensor’s ability to correctly detect environmental states or events. • Object(s) of Adaptation: The robot’s perception and state estimation processes, including sensor interpretation, filtering, and decision-making mechanisms that rely on sensed data. • Example: A robot relying on LiDAR for obstacle detection may generate false positives due to sensor noise or fail to detect transparent obstacles, which results in state estimation unpredictability and degraded navigation performance [ 41 , 42 ] .
Unavailability• Definition: Perception unavailability refers to the unavailability of perception sources, such as cameras, sensors, and LiDAR, to maintain situational awareness and operational performance [ 43 ] . • Trigger: Triggered by loss, failure, or temporary unavailability of one or more perception sources that normally provide information required for the system itself/environment perception, localization, or mapping. • Object(s) of Adaptation: The robot’s perception and sensing pipeline, including alternative sensing strategies, state estimation mechanisms, or operational behaviors that compensate for missing perception inputs. • Example: When cameras stop providing usable data due to sensor unavailability, the robot cannot continue to update its map or localize itself, leading to increased unpredictability in environment estimates and degraded operational performance [ 43 ] .
TABLE II: Perception-driven adaptation needs.
Non-determinism• Definition: Non-determinism in actuation refers to non-deterministic behavior of the actuators, where commanded actions do not always produce the predicted outcomes. Mechanical imperfections and physical effects can introduce variability in how actions translate into real-world effects, which may change the world in unpredictable ways [ 7 , 44 , 45 ] . • Trigger: Triggered by observed deviations between commanded actuator actions and the actual physical outcomes produced by the robot’s actuators. • Object(s) of Adaptation: The robot’s control strategies, motion execution models, and action outcome predictions that depend on actuator behavior. • Example: When the controller of a robotic arm commands a small change in joint angle, the actual motion may vary because of backlash, stiction, and variable friction inside the gears and bearings [ 46 ] . This variability can cause the arm to reach slightly different positions than intended.
Unavailability• Definition: Actuation unavailability refers to the unavailability or failure of actuators, which prevents the robot from executing commanded actions as intended and may compromise its ability to perform tasks [ 47 ] . • Trigger: Triggered by the detection of an actuator failure or loss of actuator functionality, resulting in the robot’s inability to execute planned motions or actions as expected. • Object(s) of Adaptation: The robot’s control strategies, motion planning, and task execution mechanisms that rely on the availability and functionality of actuators. • Example: In a legged robot, the failure of an actuator may affect the execution of a commanded motion. In this case, the failed leg moves passively while the robot’s body undergoes unintended displacement, leading to uncertainty in task execution and requiring adaptation in motion strategy or gait [ 48 ] .
TABLE III: Actuation-driven adaptation needs.
Component failure• Definition: Component failure refers to a condition in which one or more components of the robotic system no longer function as expected. Failures may be identified through sensors, external components, stakeholder input, supporting tools, supervision, or observed abnormal system behavior [ 49 , 50 , 51 , 52 ] . • Trigger : Triggered by the detection or diagnosis of abnormal behavior or system faults through monitoring, sensing, or analysis of the robot’s performance and interactions. • Object(s) of Adaptation: The robot’s operational configuration, task allocation, coordination mechanisms, or control strategies that must be adjusted to handle the detected fault and maintain system functionality. • Example: In a distributed multi-robot system, a robot suddenly starts behaving abnormally due to an internal fault; then the rest of the robots detect the inconsistency in the robot’s actions through monitoring, so the system isolates the faulty robot and reconfigures the team [ 53 ] .
Changing resources• Definition: Changing resources in the system refers to changing, degrading, failing, or fluctuating available resources in the robotic system, such as power constraints, sensors, actuators, software, or hardware [ 54 ] . • Trigger : Triggered by the detection of reduced availability, degradation, failure, or fluctuation of system resources that are necessary for executing tasks or maintaining system performance. • Object(s) of Adaptation: The robot’s resource management strategies, task allocation mechanisms, and operational planning that depend on the availability and condition of system resources. • Example: In multi-robot task allocation for long-duration missions, robots may experience component failures or environmental conditions that make certain capabilities unavailable [ 55 ] .
New resources• Definition: New resources in the system refer to the availability of new resources in the robotic system, such as power constraints, sensors, actuators, software, and hardware [ 54 ] . • Trigger : Triggered by the detection or integration of new resources or components during runtime that were not originally part of the robot’s operational configuration. • Object(s) of Adaptation: The robot’s software architecture, perception pipelines, resource management mechanisms, and task planning processes that must incorporate and exploit the newly available resources. • Example: When a newly added sensor becomes available at runtime, the robotic software architecture must adapt to integrate the new data stream [ 56 ] .
Communication impairment• Definition: Communication impairment in the system refers to communication issues, such as latency, packet loss, bandwidth limitations, interference, intermittent connectivity, synchronization errors, middleware failures, and cybersecurity threats, which constrain/degrade coordination/perception sharing in multi-robot systems or shared robotic systems with different stakeholders, and the system’s ability to operate properly [ 57 ] . • Trigger : Triggered by the detection of degraded, delayed, disrupted, or insecure communication affecting the transmission or synchronization of data among system components or collaborating robots. • Object(s) of Adaptation: The robot’s communication mechanisms, coordination strategies, data sharing protocols, and decision-making processes that depend on reliable information exchange. • Example: During drone inspection, high latency and packet loss in the wireless communication link can result in delayed or incomplete video transmission. This degraded communication can potentially compromise safe navigation and real-time decision-making [ 58 ] .
TABLE IV: System-driven adaptation needs.
Uncertain specification• Definition: Uncertain mission specification refers to the lack of complete and precise specifications of the mission objectives and the quality of performance. This may arise from abstraction, incompleteness, and inaccuracies in the formulation of mission goals and qualitative criteria. Moreover, relationships may span across single missions/tasks and multiple missions/tasks, including those with the same hierarchical level with different priorities, as well as between global and local objectives [ 20 , 24 , 59 ] . • Trigger : Triggered by encountering ambiguous, underspecified, incomplete, or conflicting mission objectives or performance criteria either before or during task execution, which prevents the robot from determining a single precise course of action. • Object(s) of Adaptation: The robot’s mission specification, including its goal, priorities, constraints, together with goal interpretation, planning strategies, and decision-making mechanisms that translate high-level mission objectives and qualitative requirements into concrete operational behaviors. • Example: A robot receives an order to “navigate efficiently and safely” without a precise specification of what efficient or safe means in measurable terms (e.g., time, energy, distance to obstacle). Since these criteria are not explicitly defined, the robot must adapt to interpret and balance these objectives when executing navigation actions [ 60 ] .
Future mission changes• Definition: Future mission changes refer to potential changes (addition, removal, or update) during the system’s operation in the tasks of the mission or in the overall mission objectives defined at design time [ 61 , 62 , 59 , 63 ] . • Trigger : Triggered by the detection of modifications to mission definitions, such as the addition, removal, or update of missions or subtasks during system operation. • Object(s) of Adaptation: The robot’s mission planning, task decomposition, scheduling, and coordination mechanisms that manage the execution of missions and their subtasks. • Example: During runtime, a robotic system may receive an updated mission in which a new subtask is created or an existing one is removed. The change can affect other missions executed in the shared environment and cause inconsistency in all missions [ 64 , 61 ] .
Outdated mission• Definition: Outdated mission refers to the condition in which a mission no longer needs to be executed because its objectives have already been achieved, are impossible to achieve, or are no longer required [ 59 ] . • Trigger : Triggered by the detection that a mission objective or performance requirement is no longer applicable, already satisfied, or cannot be meaningfully pursued during system operation. • Object(s) of Adaptation: The robot’s mission planning, goal management, and decision-making processes that determine which objectives to pursue and how to allocate effort. • Example: In a multi-robot system, a robot receives a task to inspect a specific area. If the task has already been satisfied by another robot’s presence in the environment, the original objective becomes redundant.
Uncertain quality goals• Definition: Uncertain quality goals refer to failures, changes, conflicts, and additions of non-functional goals in robotic missions. These goals, such as safety, reliability, robustness, efficiency, security, and real-time performance, are sensitive to environmental dynamics, mission interdependencies, and shared resource consumption [ 65 , 66 ] . • Trigger : Triggered by the detection of changes, conflicts, degradation, or new additions in non-functional goals that affect how missions should be executed. • Object(s) of Adaptation: The robot’s decision-making, planning, and control strategies that balance and enforce non-functional requirements during task execution. • Example: A hospital - service robot may need to adapt when its planned path becomes blocked by a human worker or another robot. It must remain robust in low - light conditions, compute efficient paths to save time and energy, and ensure safe physical interaction during manipulation tasks [ 67 ] .
Uncertain relationships• Definition: Uncertain relationships refer to conflicting, implicit, unknown, or unclear dependencies and interactions among missions at a given time, which are hard to capture in a deterministic way. Missions may share the same physical environment where robots operate. Execution of a mission may cause inconsistencies and make other missions unreachable/uncomplishable, and have an impact on qualitative aspects. Additionally, missions may be executed in parallel or sequentially, which may share, consume, and occupy the same environmental and robotic resources. So, the execution of one mission can affect another mission [ 68 , 69 , 70 ] . • Trigger : Triggered by the detection of conflicts, dependencies, or unintended interactions between missions, particularly when their execution affects shared resources, environment states, or each other’s feasibility. • Object(s) of Adaptation: The robot’s mission coordination, scheduling, and planning mechanisms that manage dependencies, conflicts, and interactions between multiple missions. • Example: In a shared environment, one robot may move an object as part of a mission, while another robot depends on the object being in the original location for a different task. This interaction can degrade the performance of the second mission [ 71 ] .
TABLE V: Mission-driven adaptation needs.
Execution context changes• Definition: Execution context changes refer to changes in the robot operational environment with which the robotic system interacts. The context changes may arise either in the form of static (e.g., a healthcare robot suddenly transferred to a different domain) or dynamic (e.g., a rover faces a new type of terrain). Due to ongoing changes beyond the control of the developer and robotic system, the context could not be known during system design [ 59 , 72 , 73 , 74 ] . • Trigger : Triggered by the detection of significant and unforeseen changes in the operating context or system structure that were not anticipated during design time. • Object(s) of Adaptation: The robot’s system architecture, behavior, mission, perception, actuation, knowledge of the environment, and interaction strategies that must adjust to operate effectively under new or significantly altered conditions. • Example: A rover on Mars may encounter a completely new type of terrain due to wind effects on Mars. Because the environmental conditions differ from prior assumptions [ 20 ] .
Non-human agents coex.• Definition: Non-human agents coexistence refers to the presence of other non-human entities, such as robots, autonomous systems, or animals, whose behavior may be unpredictable, difficult to model, and/or capable of changing the environment in ways that affect the robot’s mission [ 75 ] . • Trigger : Triggered by the detection of changes in the environment or task conditions caused by the actions of other non-human agents, especially when their behavior deviates from expected or modeled patterns. • Object(s) of Adaptation: The robot’s perception, planning, and interaction strategies used to operate safely and effectively in environments shared with other autonomous or semi-autonomous agents. • Example: In a shared workstation, another robot may unintentionally change the location of an object that must be picked by a cobot.
Human agents coex.• Definition: Human agents coexistence refers to the presence of humans whose behavior, intentions, preferences, or interactions may be unpredictable, difficult to model, and/or capable of changing the environment, task conditions, or mission requirements in ways that affect the robot’s mission [ 76 , 77 , 78 ] . It also introduces safety-critical constraints, as the robot may need to adapt its behavior to avoid harmful interactions, maintain safe distances, and operate in a socially acceptable manner. • Trigger : Triggered by the detection of human presence, behavior changes, or interactions that deviate from expected patterns or introduce uncertainty, safety concerns, or ethical constraints. • Object(s) of Adaptation: The robot’s perception, planning, interaction, and decision-making strategies, particularly those related to human-aware behavior and ethical considerations. • Example: A robot operating in the environment in the presence of a human may need to adapt its behavior to secure human privacy. The robot must avoid capturing sensitive visual data or modify its sensing and navigation strategies when people are nearby.
TABLE VI: Environment-driven adaptation needs.
Fig. 3: Respondent’s perception of the relevance of Adaptation Needs in robotics.
Limit.Description
L1Static control flow in classical BTs limits runtime flexibility [ 17 , 16 ] .
L2Fixed execution order prevents flexible adaptation to runtime changes [ 16 , 82 ] .
L3Classical BTs often operate under deterministic assumptions, which are unsuitable for uncertain environments [ 82 , 20 , 83 , 84 ] .
L4Difficulty handling partial observability and incomplete state information [ 18 , 17 ] .
L5Human-centered and collaborative industrial environments introduce uncertainty and variability not captured by classical BTs [ 84 , 82 ] .
L6Suboptimal execution and lack of explicit representation of the purpose and intent of behaviors [ 16 ] .
TABLE VII: Limitations of adaptability and extensibility of BTs.
 SubtypeModularityInterpretabilityReactivityReusabilityMaintainabilityExtensibilityDecomposabilityFlexibilityBT limit.
KnowledgeAbstraction        -
 Incompleteness✓\checkmark ∼\sim✓\checkmark✓\checkmark✓\checkmark✓\checkmark L4
 Inaccuracy  ✓\checkmark     -
 Drift✓\checkmark✓\checkmark✓\checkmark✓\checkmark✓\checkmark  ✓\checkmark-
Percep.Multiplicity        -
 Inaccuracy        -
TABLE VIII: Mapping adaptation needs to BT features and limitations. Legend: ✓\checkmark sufficient; ∼\sim partial/implicit; ×\times no support; □\square no information.
Fig. 4: Practitioner responses on how each family of adaptation needs should be addressed using BTs.
 SubtypeEvol.Exten.Gen.Refin.
Knowled.Abstraction    
 Incompleteness✓\checkmark ( [ 88 ] )✓\checkmark ( [ 18 ] )∼\sim ( [ 89 , 90 ] )✓\checkmark ( [ 90 ] )
 Inaccuracy  ∼\sim ( [ 90 ] )✓\checkmark ( [ 90 ] )
 Drift✓\checkmark ( [ 91 ] ) ∼\sim ( [ 92 ] ) 
Percep.Multiplicity    
 Inaccuracy    
TABLE IX: Mapping adaptation needs to BT adaptation and evolution approaches. Legend: ✓\checkmark suitable; ∼\sim partially suitable; ×\times not suitable; □\square no information.
Fig. 5: Mean suitability of four BT adaptation approaches across five families of adaptation needs, as rated by 46 practitioners on a three-point scale (1 = Not suitable, 2 = Partially suitable, 3 = Completely suitable). Each panel shows one approach; the polygon shape reveals its suitability profile across families.
StudyYearMain focusDifference from our work
Iovino et al. [ 10 ]2022Survey of BTs in robotics and game AI, including concepts, applications, design patterns, synthesis, and learning.Provides a broad overview of BTs, but does not organize the analysis around robotic adaptation needs or assess classical BT limitations with respect to such needs.
Ögren and Sprague [ 11 ]2022Review of BTs from a robot-control perspective, including modularity, hierarchy, reactivity, planning, and control.Focuses on BT principles and control-oriented properties rather than on adaptation needs and adaptation-oriented BT techniques.
Biggar et al. [ 12 ]2020Analysis of BTs as a model of action selection, with emphasis on modularity, reactivity, and formal properties.Studies BTs as an action-selection formalism, but does not provide a robotics-specific taxonomy of adaptation needs or map such needs to BT capabilities.
Shin and Jung [ 21 ]2024Survey of BT-based task planning for autonomous robotic systems, especially reinforcement learning and learning from demonstration.Focuses on learning-based task planning, whereas our work covers broader adaptation needs and different forms of BT enhancement.
Ghzouli et al. [ 13 ]2023Empirical study of BTs and state machines in ROS-based robotics applications.Studies language concepts, usage patterns, and reuse in robotic software, but does not analyze adaptation needs or classify BT adaptation mechanisms.
Iovino et al. [ 105 ]2023Comparison of the programming effort required to implement and modify BTs and finite state machines.Focuses on implementation and modification effort, but does not investigate why robotic systems need adaptation or how BT-based approaches address such needs.
TABLE X: Positioning of our work with respect to related studies.
More in this topic

Behavior Trees

Behavior trees are hierarchical, reactive control structures for organizing tasks, actions, and recovery in robots and autonomous agents.

See all 19 papers

Explore similar work

CardsList
Sep 18, 2026

Behavior Trees for Robotic Systems: An Empirical Study on Practices and Experiences

Over the last decade, behavior trees (BT) have become one of the dominant behavior models for coordinating missions of robotic systems. Yet empirical evidence on BT adoption in real-world contexts remains limited, especially regarding practitioners experiences and practices. Without practitioner-grounded evidence from both academic and industrial settings, the research community risks developing guidelines and tools that are plausible in principle but only partly aligned with real challenges. This scarcity of evidence leads to ad-hoc practices, impeding software reuse, maintenance, and evolution. This paper reports a mixed-methods study combining a technical action research investigation at an automotive company with a survey of 34 robotics practitioners. Our results indicate that BTs improve team communication and the understandability of robotic decision-making logic, reflecting BTs practical value beyond mission coordination. At the same time, practitioners face multiple non-trivial design and integration decisions, complicated by the lack of adequate guidelines and tool support. These decisions span architectural and language choices when implementing BTs and integrating them within ROS, for which we report observed patterns. Additionally, practitioners faced multi-factor granularity decisions and mixed experiences with current libraries. For the most used BT libraries, they reported implementation challenges and documentation gaps. Regarding planning algorithms, practitioners find deciding optimal BT node ordering challenging. Views on scalability remained inconclusive, and limitations in current libraries and practices hinder broader adoption. We conclude with cross-cutting observations and implications for both practitioners adopting BTs in robotic systems and researchers aiming to advance empirical understanding of BT adoption.

Razan Ghzouli, Jennifer Horkoff, Daniel Strüber +1

Department of Computer Science and Engineering, Chalmers University of Technology and University of Gothenburg, Gothenburg, Sweden · Department of Software Science, Radboud University, Nijmegen, The Netherlands · Software and Societal Systems Department, Carnegie Mellon University, Pittsburgh, PA, USA

Behavior TreesRoboticsRobot Task Planning
AskPDF
Mar 17, 2026

CABTO: Context-Aware Behavior Tree Grounding for Robot Manipulation

Behavior Trees (BTs) offer a powerful paradigm for designing modular and reactive robot controllers. BT planning, an emerging field, provides theoretical guarantees for the automated generation of reliable BTs. However, BT planning typically assumes that a well-designed BT system is already grounded -- comprising high-level action models and low-level control policies -- which often requires extensive expert knowledge and manual effort. In this paper, we formalize the BT Grounding problem: the automated construction of a complete and consistent BT system. We analyze its complexity and introduce CABTO (Context-Aware Behavior Tree grOunding), the first framework to efficiently solve this challenge. CABTO leverages pre-trained Large Models (LMs) to heuristically search the space of action models and control policies, guided by contextual feedback from BT planners and environmental observations. Experiments spanning seven task sets across three distinct robotic manipulation scenarios demonstrate CABTO's effectiveness and efficiency in generating complete and consistent behavior tree systems.

Yishuai Cai, Xinglin Chen, Yunxin Mao +6

National University of Defense Technology · PsiBot · PKU-Psibot Lab +1

Robotic ControlRobotic ManipulationRobot Task PlanningBehavior Trees
AskPDF
Oct 18, 2026

Self-Crystallizing Behavior Trees: Amortizing Foundation-Model Reasoning into Deterministic Control for Autonomous Drones

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±1.411\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±2%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.

Behnam Moradi

Software Department, KavAI. California, USA

Behavior TreesRobot Policy AdaptationLanguage Model-Based ControlAerial Robotics
AskPDF