Task Composition Method for Behavior-Based Control of Uncrewed Surface Vehicles

Article information

J. Ocean Eng. Technol. 2026;40(4):522-530
Publication date (electronic) : 2026 July 24
doi : https://doi.org/10.26748/KSOE.2026.030
1Researcher, Agency for Defense Development, Daejeon, Republic of Korea
Corresponding author Uijong Kim: uijongkim@add.re.kr
It is noted that this paper is a revised edition based on the author’s master’s thesis submitted to KAIST (Korea Advanced Institute of Science and Technology) (Kim, 2023).
Received 2026 May 4; Revised 2026 June 21; Accepted 2026 June 27.

Abstract

As the autonomy of uncrewed systems advances, there is increasing demand for robust operation in complex environments. In the maritime domain, the dynamic and unstructured nature of the ocean poses significant challenges for uncrewed surface vehicles (USVs) to respond in real time using conventional optimization-based methods. This paper proposes a behavior-based control framework built from simple, modular behaviors. Each behavior is formulated as a real-valued objective function over a shared action space, and the behaviors are combined through task composition using lattice-algebraic operations–disjunction, conjunction, and negation. This approach yields a scalable arbitration scheme that integrates new behaviors by accounting for their logical interrelationships. The performance of the framework was validated through 1,000 Monte Carlo trials of an integrated multi-agent scenario combining formation maintenance, multi-target rescue, and static obstacle avoidance. Compared to a baseline that aggregates the same behaviors by a weighted sum, the proposed method reduces the root-mean-square (RMS) position error from 31.63 m to 22.78 m, raises the rescue rate from 81.74% to 90.32%, and reduces the collision rate from 4.56% to 0.85%, with non-overlapping 95% confidence intervals on every metric. The proposed architecture offers a reliable and scalable control solution for multi-objective USV operations.

1. Introduction

The recent advances in uncrewed mobility technologies have led to an increasing demand for sophisticated robot system architectures capable of addressing complex operational environments (Khatib, 1986; Kortenkamp et al., 2016). Traditional optimization-based control can guarantee optimality but often becomes computationally prohibitive and difficult to scale in complex environments. Brooks (1986) proposed behavior-based robotics to address these limitations, in which the controller is decomposed into a set of simple, concurrently running behavior modules. Within the behavior-based paradigm, arbitration algorithms can be broadly classified into competitive and cooperative schemes (Arkin, 1989; 1998). A competitive scheme selects the output of a single behavior as the system response, whereas a cooperative scheme combines the multiple behaviors simultaneously into one composite action. This study adopts the cooperative scheme rather than a predefined sequence of actions because operating missions in unstructured environments are addressed by concurrently executing multiple non-sequential behaviors. Fig. 1 contrasts the two paradigms: the traditional deliberative architecture, in which sensing, planning, and acting are executed sequentially (Fig. 1(a)), and the behavior-based architecture, in which perception data are processed by parallel, concurrently running behaviors (Fig. 1(b)).

Fig. 1

Comparison of robot control architectures: (a) Traditional deliberative architecture based on a sequential sense-plan-act pipeline, and (b) the proposed behavior-based control architecture, which processes perception data through parallel, concurrent behaviors for reactive responses

The maritime domain places particularly demanding requirements on autonomous control. Vessels must continuously perceive and react to unstructured environments while performing multiple missions such as reconnaissance, detection, and rescue operations under intermittent communication and tight real-time constraints. These conditions make scalable, modular, and reactive autonomy a critical research priority for uncrewed surface vehicles (USVs). Frameworks such as MOOS-IvP (Benjamin, 2002; Benjamin et al., 2006; Benjamin et al., 2009; Benjamin et al., 2010) and operator-aware behavior-based USV controllers (Cho et al., 2019) have been developed to address this need. CARACaS (Elkins et al., 2010; Wolf et al., 2017) has further validated the approach through at-sea trials of autonomous navigation and cooperative operation.

Although these studies have shown the potential of behavior-based control in maritime environments, a key limitation remains at the arbitration layer. Existing cooperative schemes resolve conflicts among simultaneously active behaviors by arithmetic aggregation of their objective functions—most prominently the weighted-sum formulation used by MOOS-IvP, in which each behavior contributes a piecewise-linear objective function on a shared decision space and the solver selects the action that maximizes the weighted combination of these functions via branch-and-bound. While more expressive than subsumption or potential-field averaging, such utility aggregation collapses the active behaviors into a single composite scalar. Therefore, when two behaviors disagree, the composite peak is pulled toward the weighted mean of their preferred actions—an effect referred to as the averaging problem.

Accordingly, this paper proposes a task-composition method that coordinates behaviors through their logical interrelationships rather than through scalar weighting. Building on prior formulations from multi-objective action selection (Pirjanian, 1999; Van der Pol & Oliehoek, 2014), value-function composition (Van Niekerk et al., 2019; Sutton & Barto, 1998), and Boolean task algebra (Nangue Tasse et al., 2020), each behavior is formalized as an explicit objective function on a shared action space, and behaviors are combined by lattice-algebraic operators: pointwise disjunction, conjunction, and negation defined as the pointwise maximum, minimum, and Boolean complement, respectively. This shift in arbitration paradigm yields two operational departures from utility aggregation. First, the multimodal structure of the composite objective is preserved by construction, so disagreeing behaviors are not collapsed into a single averaged peak. Second, adding a new behavior requires only the specification of its logical relation to the existing set, eliminating the per-behavior

The primary contributions of this research are as follows: (1) design of a USV-specific behavior-based control framework, an architecture that incorporates USV dynamics to design primitive behaviors for autonomous maritime operations; (2) development of logical arbitration via Task Algebra, logical operations used to define behavior interactions, ensuring scalability and systematic mission composition; (3) verification of multi-objective optimization: Each behavior is formalized in an action space as an objective function. Superior performance was demonstrated in integrated scenarios that are difficult to resolve with arithmetic methods.

Fig. 2

Hierarchical structure of task composition. The overarching mission is logically composed of multiple independent subtasks, where each sub-task represents a specific behavior encoded as an objective function. scalar-weight tuning step required whenever the behavior set is modified. Fig. 2 presents the resulting hierarchical structure of task composition, in which the overarching mission is logically composed of multiple independent sub-tasks, each corresponding to a behavior encoded as an objective function.

2. Behavior-based System for USV

A baseline behavior-based control architecture was designed to implement the behavior-based framework for USVs and their mission operation (Fig. 3). Following the input flow, the architecture comprises three main components: a behavior manager, a parallel bank of behaviors, and a behavior arbitrator. Events or Perception information from the perception/sensing system reach the behavior manager, where primitive behaviors, such as waypoint tracking and obstacle avoidance, run in parallel and each produces an individual objective function over the action space. The behavior arbitrator then composes these objective functions and emits a single guidance signal, which the controller realizes via PD rudder control and feed-forward thrust to drive the USV dynamics as described in Fossen (2011), closing the perception–action loop. The behavior-based framework for USV adopted in this study was built on Cho et al. (2019).

Fig. 3

Behavior-based control architecture for USV. Environmental perception triggers a parallel bank of behaviors within the behavior manager. The behavior arbitrator combines their respective objective functions to generate an optimal guidance signal, which is executed by controllers driving the USV dynamics.

The action space was defined as the set of target motion states reachable from the current motion state and is expressed as ordered pairs of target speed and target course according to the following equation:

(1) A={〈ψi, Vi〉∣   ∣Vi-V∣〈rV,V〉 0,∣ψi-ψ∣〈γψ}

where Vi and ψi denote elements of the action-space set A, while V (a) Traditional control architecture (b) Behavior-based control architecture and ψ denote the current speed and course of the vehicle, respectively. The threshold values γV and γψ are imposed on the admissible increments in speed and course to restrict the action space to target speeds and courses that are reachable from the current state, as defined in Eq. (1).

For reference, the conventional approach, hereafter referred to as the Linear Composition, selects the optimal action as the argmax of a weighted sum of individual objective functions, where the objective function for each behavior is defined as fk over the action space, and the optimal action is given by the following equation:

(2) 〈ψ^^, V^〉=argmax〈ψi,         Vi〉 ∈A   [Σk=1n wkfk(ψi,         Vi)]

where 〈ψ̂, V̂〉 denotes the ordered pair representing the optimal action, and wk denotes the weight assigned to the kth behavior. Behaviors with higher priority for an uncrewed vehicle to perform are assigned correspondingly higher weights. While straightforward to implement, this scheme cannot capture the logical interrelationships among simultaneously active behaviors. The proposed Task Composition method, described in Section 3, addresses this limitation by replacing this weighted summation with lattice-algebraic operators.

2.1 Waypoint Tracking Behavior

The waypoint-tracking behavior steers the vehicle toward a sequence of waypoints, each with inner/outer radii defining when it is reached and the next is engaged. The desired heading and speed are obtained from guidance laws, and the objective function is defined from heading and speed potentials:

(3) OGO(ψi,Vi)=1-αψ|ψi-ψd|+αV|Vi-Vd|max〈ψj,Vj〉 ∈A[αψ|ψj-ψd|+αV|Vj-Vd|]

where, (ψi, Vi) is a candidate action in the discretized action space A. (ψd, Vd) is the desired heading and speed obtained from the guidance law, and αψ, αV > 0 weight the heading and speed errors. The denominator normalizes the weighted error against the worst case over the action space.

2.2 Rescue Behavior

The rescue behavior maneuvers the vehicle toward a point-of-interest (drowning person) whose position is provided by the perception system. Each target is treated as a goal point: the rescue objective takes the same action-space form as the Go-behavior objective in Eq. (3). By contrast, the waypoint-tracking behavior is associated with a single waypoint at any given moment; the rescue behavior can be instantiated for multiple targets simultaneously. The guidance algorithm, however, differs from that of waypoint tracking; rather than following a stored path, the rescue guidance recalculates the approach command at each control step from the current target position and own-ship state.

2.3 Formation Control Behavior

The formation behavior steers each agent toward a desired position defined relative to the formation pattern (Balch & Arkin, 1998). The guidance command toward this desired position is calculated by decomposing the position error into two orthogonal components. (Fig. 4) The fore-aft error is regulated by proportional control on the desired speed Vd, while the lateral error is regulated by proportional control on the desired heading ψd. Three coaxial zones around the desired position gate the proportional gains: dead zone, controlled zone, and ballistic zone. The decomposed pair (ψd,Vd) produced by this guidance step is then taken as the desired action of the formation behavior and is fed into the go-behavior objective of Eq. (3). The desired speed is set to a nominal cruise speed Vnominal, modulated by a zone-dependent proportional gain Kp (smaller in the dead zone, larger in the ballistic zone outside the controlled zone); Table 1 lists the specific values.

Fig. 4

Mechanism of the formation control behavior. The desired position for the formation is derived using an average-based method (Balch & Arkin, 1998). The position error relative to the desired formation point is decoupled into a fore-aft error regulated by speed control, and a lateral error regulated by heading control.

Simulation parameters.

2.4 Collision Avoidance Behavior

The collision-avoidance behavior treats each obstacle as a two-dimensional particle of given radius and is formulated in terms of the distance to the closest point of approach (DCPA) between the candidate own-ship action and the obstacle. For a candidate action (ψi,Vi) in the action space A, the objective function is expressed as

(4) Oavd(ψi,Vi)={0if DCPA(i)<γminDCPA(i)-γminγmax-γminif γmin≤DCPA(i)≤γmax1if DCPA(i)>γmax

where γmin and γmax are the lower-bound threshold and higher-bound threshold for the minimum-safe and no-risk DCPA distances.

3. Task Composition

Task algebra provides an analytical framework for the logical composition of behaviors. Building on the value-function composition of Van Niekerk et al. (2019), Schaul et al. (2015) and the Boolean task algebra of Nangue Tasse et al. (2020), this formulation defines pointwise disjunction (∨, supremum), conjunction (∧, infimum), and negation (¬) over real-valued objective functions, with the negation supplied by the standard De Morgan and Boolean structure (Gratzer, 2002; 2011). The formal definitions and properties used in this paper are summarized in the Appendix. The disjunction and conjunction over tasks reduce to pointwise extrema of their objective functions:

(5) OMa ∨Mb(a)=max(OMa(a),   OMb(a))
(6) OMa ∧Mb(a)=min(OMa(a),   OMb(a))

Negation is the Boolean complement with respect to two distinguished universal bounds: a universal task MU whose objective function OU(a) attains the maximum admissible value at every action. An empty task Mø whose objective function Oø(a) attains the minimum value at every action. The negation of a task M is then defined pointwise on the action space as follows:

(7) O¬M (a)=(OU(a)+O∅(a))-OM(a)

The composite tasks defined in Eqs. (8)–(11) are evaluated action-by-action using these extrema.

Existing USV behavior-activation schemes assume that a single event triggers each behavior and do not arbitrate among simultaneously active behaviors (e.g., multiple drowning persons). This is addressed with a task-composition layer in which each primitive behavior is identified with a sub-task and is denoted by M. The tasks considered in this work differ only in their objective functions. Hence, a pointwise comparison induces a partial order on tasks under which the disjunction (∨), conjunction (∧), and negation (¬) are defined on the set of sub-tasks as in Eqs. (5)–(7), with M serving as their operand. New behaviors can be added without retuning the scalar weights; only their logical relation to the existing set needs to be specified.

Fig. 5 shows the three algebraic operators on a two-dimensional action space A. Each panel plots the objective value of a task, with brighter regions indicating preferred actions. Disjunction (M1∨M2) takes the pointwise maximum. Hence, any action that satisfies either base task remains a high-value option—the desired behavior for “reach waypoint” or “rescue any of several targets.” Conjunction (M1∧M2) takes the pointwise minimum, so an action is acceptable only when it satisfies every sub-objective simultaneously: the desired behavior for “avoid every obstacle while tracking the waypoint.” Negation (¬M1 = 1 − M1) inverts the value map, turning attractors into repellers, which is useful for restricted-zone avoidance. Crucially, all three are pure pointwise operations on the objective functions; new behaviors are added simply by extending the algebraic expression.

Fig. 5

Operator semantics over the action space: (left to right) two base tasks M1, M2, the disjunction M1∨M2 (pointwise max), the conjunction M1∧M2 (pointwise min), and the negation ¬M1 = 1 − M1. Brighter regions correspond to higher objective values; the star marker denotes the action selected by argmax over each composite map.

Fig. 6 contrasts the algebraic disjunction (∨) with a Linear Composition on a two-target rescue scenario. Two rescue objectives MR1 and MR2 are defined for targets located at ψd1 = −25° and ψd2 = +25°, respectively, each producing a unimodal value map peaked at its target bearing (a) and (b). (c) shows the disjunction MR1∨MR2. Both peaks are preserved because the operator returns the pointwise maximum, and the argmax (white star) selects the action that drives the own-ship toward one of the targets. When several rescue behaviors are simultaneously active, and the disjunction admits more than one argmax, the tie is resolved by selecting the action with the smallest target index. Panel (d) shows the Linear Composition (0.5 MR1 + 0.5 MR2 used by Linear Composition): the two peaks merge into a single peak halfway between the targets, and the argmax instructs the own-ship to head straight ahead, reaching neither target. This well-known failure of Linear Composition is naturally avoided by the algebraic disjunction, which preserves the multimodal structure of the composite objective without requiring weight tuning.

Fig. 6

Comparison between Task Composition (algebraic disjunction) and Linear Composition (weighted summation) in a multi-target rescue scenario. (a, b) Individual rescue objectives (MR1, MR2); (c) Disjunction preserves the multimodality of the action space (MR1∨MR2), allowing commitment to a specific target; (d) Linear Composition improperly averages the objectives (0.5 MR1 + 0.5 MR2), resulting in an invalid action straight between the targets.

The conjunction (∧) operator plays the complementary role; it is the right choice when several constraints must be satisfied simultaneously. Fig. 7 presents this in a collision avoidance situation where the waypoint-tracking behavior operates simultaneously with collision avoidance. Panel (a) shows the baseline waypoint-tracking objective Mwpt, peaked at the desired course ψ = 0° and speed u=4 m/s. Panel (b) shows the collision avoidance objective Mavoidance with a continuous linear gradient in between. Hence, the unsafe region forms a wedge centered on the bearing to the obstacle. The conjunction Mwpt∧ Mavoidance in panel (c) collapses to a narrow region on the side where both sub-objectives are simultaneously satisfied.

Fig. 7

Action-space conjunction for simultaneous waypoint tracking and obstacle avoidance. (a) Waypoint tracking objective (Mwpt); (b) Collision avoidance objective (Mavoidance); (c) Conjunction (Mwpt ∧ Mavoidance) steers a safe evasive maneuver.

A mission can be expressed through logical operations between individual behaviors, e.g., by combining waypoint tracking, rescue maneuvers, and collision avoidance simultaneously.

Throughout this section, let R and S denote the set of currently visible rescue targets and the set of static obstacles within the perception range, respectively; their cardinalities vary over time, and the indices r and s in the composite tasks below range over R and S, respectively. This framework allows for the seamless integration of new behaviors by simply defining their logical relationships with existing ones. In this study, the final composite task arbitrates the ‘GO’ and ‘Avoid’ groups through a conjunction operation.

(8) MRescue=∨r∈RMrescue
(9) MGO=Mwpt∨∨r∈RMrescue
(10) MAvoid=∧obs∈SMaviodobs
(11) MFin=MGO∧MAviod

4. Simulation and Discussion

The proposed procedure for automatic navigation using behavior-based control was validated through 1,000 Monte Carlo simulations in a USV simulator. Two action-selection schemes were compared under identical scenarios. The action space is the set of feasible (target speed, target course) pairs reachable from the current state, and each scheme selects one element of this set at every control step.

  • (1) Linear Composition: each behavior exposes its objective function on the action space, and the final action is the argmax of a weighted sum of those objectives, as expressed in Eq. (2).

  • (2) The proposed Task Composition: the action is selected by lattice-algebraic composition of the objective functions, returning the argmax of the composite task MFin defined in Eq. (11).

4.1 Simulation Setup

All Monte Carlo trials were run in a MATLAB simulator implementing the USV dynamics. Within each trial, static obstacles and drowning persons are generated randomly within the prescribed regions; the trial terminates when the formation reaches its waypoint within the acceptance radius or when the episode time limit is exceeded. Table 1 lists the complete set of simulation parameters used throughout this section.

The Linear Composition baseline implements the cooperative action-selection method reported by Cho et al. (2019), in which behavior-wise utility functions on a shared action space are combined by a weighted sum and the optimal action is taken as the argmax over the action space (Eq. (2/0)). Table 2 lists the per-behavior weights wk used in the baseline. These weights were selected so that each behavior asserts an appropriate level of influence at its critical operating point, intentionally placing the baseline at a favorable operating regime.

Behavior weights used in the Linear Composition (baseline).

4.2 Simulation Results

Behavior arbitration methods were applied to a multi-agent scenario in which four agents maintained a horizontal formation while rescuing drowning people and avoiding collisions. Three performance metrics are reported. The position error is the root-mean-square (RMS) Euclidean distance between the actual position of each agent and its desired position over all agents and over the duration of the episode. The rescue rate is the percentage of trials in which all rescue targets were successfully reached within the episode. The collision rate is the percentage of trials in which at least one inter-vessel distance falls below a predefined collision radius at any time during the trial. All three metrics are reported as the mean value followed by the 95% confidence interval in brackets across the 1,000 trials; the interval is calculated using the Wilson-score method for the two Bernoulli rate metrics and by a normal approximation using the sample standard deviation for the Position error. Table 3 lists the results of the integrated scenario, and Fig. 8 presents a representative trial as a time-lapse.

Results of the integrated scenario (1,000 trials)

Fig. 8

Time-lapse visualization of the integrated multi-agent scenario. Four USVs successfully maintain a horizontal formation while dynamically breaking off to rescue scattered targets (green squares) and actively avoiding static obstacles (black circles), showing robust multi-objective arbitration.

Task Composition achieved a lower position error (22.78 m), a higher rescue rate (90.32%), and a lower collision rate (0.85%); the Linear Composition baseline cannot natively handle multiple simultaneous rescue targets and yields substantially poorer performance on every metric. This scenario best highlights the structural difference between the two methods, as the advantage of Task Composition over Linear Composition is most pronounced when behaviors have conflicting objectives. Although the Linear Composition baseline was given intentionally favorable weights (Table 2), it still underperformed the proposed method on every reported metric, suggesting that the observed advantage arises from the structural property of lattice operators (preservation of multimodal structure) rather than from suboptimal weight tuning of the baseline.

Linear Composition and Task Composition have a comparable per-step computational cost because both evaluate the same primitive objective functions on the same discretized action space. The only difference is whether the aggregator is a weighted sum or a pointwise maximum/minimum. The design effort, however, is shifted. The logical composition replaces real-valued weights with categorical logical relations, which are easier to specify when behaviors are added or modified because a new behavior requires only its objective function and its logical relation (∨, ∧, or ¬) to the existing set.

5. Conclusions

This paper proposed a novel behavior-based control framework for USVs designed to implement autonomy in dynamic and unpredictable maritime environments. The proposed system leveraged a logical task-composition layer to overcome the limitations of Linear Composition (conventional weighted-summation arbitration). By formulating primitive behaviors, such as waypoint tracking, obstacle avoidance, formation maintenance, and rescue, as objective functions over a shared action space, complex mission profiles were systematically assembled using lattice-algebraic operators.

The effectiveness of this framework was validated rigorously through Monte Carlo simulations involving integrated, multi-objective scenarios. The proposed Task Composition method explicitly addressed the “averaging problem” inherent in Linear Composition. Consequently, the framework showed significantly superior performance across all evaluation metrics. Compared to the Linear Composition baseline, the proposed method achieved a substantially reduced position error (22.78 m vs. 31.63 m), enhanced the multi-target rescue success rate (90.32% vs. 81.74%), and minimized the collision rate to only 0.85%. These results confirmed that algebraic arbitration showed precise mission execution and safe evasive maneuvers, even when multiple conflicting objectives arise simultaneously.

A major contribution of this architecture is its structural scalability. Integrating new behaviors into the system requires only the definition of their objective functions and their logical relationships to existing tasks. Furthermore, the framework supports optional per-behavior scaling within the logical composition, allowing operators to fine-tune the relative influence of individual behaviors. The framework will be deployed and validated on physical USV platforms through at-sea field tests to confirm its real-world applicability and reliability under practical environmental disturbances.

Notes

No potential conflict of interest relevant to this article was reported.

This study was supported by the Defense Acquisition Program Administration and the Agency for Defense Development under Grant 915057401.

Appendices

Appendix

Task Algebra

Definition 1

A Boolean algebra is a set B equipped with the binary operators ∨ (disjunction) and ∧ (conjunction), and the unary operator ¬ (negation), which satisfies the following Boolean algebra axioms for a, b, c in B:

  • (i) Idempotence: a ∧ a = a ∨ a = a.

  • (ii) Commutativity: a ∧ b = b ∧ a and a ∨ b = b ∨ a.

  • (iii) Associativity: a ∧ (b ∧ c) = (a ∧ b) ∧ c and a ∨ (b ∨ c) = (a ∨ b) ∨ c.

  • (iv) Absorption: a ∧ (a ∨ b) = a ∨ (a ∧ b) = a.

  • (v) Distributivity: a ∧ (b ∨ c) = (a ∧ b) ∨ (a ∧ c) and a ∨ (b ∧ c) = (a ∨ b) ∧ (a ∨ c).

  • (vi) Identity: there exists 0, 1 in B such that

    0∧a=00∨a=a1∧a=a1∨a=1

  • Complements: for every a in B, there exists an element a′ in B such that a ∧ a′ = 0 and a ∨ a′ = 1 .

Proof of Theorem 1

Assumption

For all tasks M ∈ ℳ, the objective function OM(a) takes values only in {O∅, OU} = {0, 1} for all a ∈ A; that is, each action either fully satisfies (OM(a) = 1) or fully fails to satisfy (OM(a) = 0) the task.

Theorem 1

Let M be a set of tasks which adhere to the Assumption. Then (M, ∨, ∧, ¬, MU, M∅) is a Boolean algebra.

Proof

Let M1, M2 ∈ M. We show that ¬, ∨, ∧ satisfy the Boolean properties (i) – (vii).

(i)–(v): These easily follow from the fact that the min and max functions satisfy the idempotent, commutative, associative, absorption and distributive laws.

(vi): Let OMU∧¬M1 (α) and OM1 (α) be the objective functions for MU (α)∧M1 (α) and M1 (α) respectively. Then for all a ∈ A,

OMU∧¬M1 (α)=min{OMU(α),OM1(a)}=OM1(a)

Thus MU ∧ M1 = M1. Similarly MU∨M1 = MU, M∅∧M1 = M∅, and M∅∨M1 = M1. Hence M∅ and MU are the universal bounds of M.

(vii): Let OM1∧¬M1 (a) be the objective function for M1∧¬M1. Then for all a ∈ A, since OU(a) = 1 and O∅(a) = 0:

OM1∧¬M1 (a)=min{OM1(a), (OU(a)+O∅(a))-OM1(a)}=min{OM1(a),1-OM1(a)}

By Assumption, OM1 (a) ∈ {0, 1} for all a ∈ A. Therefore:

  • – If OM1(a) = 1: min{1, 1−1} = min{1, 0} = 0 = O∅(a)

  • – If OM1(a) = 0: min{0, 1−0} = min{0, 1} = 0 = O∅(a)

Thus OM1∧¬M1(a) = O∅(a) for all a ∈ A, so M1 ∧ ¬M1 = M∅. Similarly M1 ∨ ¬M1 = MU.

References

Arkin R. C. 1998;Behavior-based robotics MIT Press;
Arkin R. C. 1989;Motor schema-based mobile robot navigation. The International Journal of Robotics Research 8(4):92–112. https://doi.org/10.1177/027836498900800406.
Balch T, Arkin R. C. 1998;Behavior-based formation control for multirobot teams. IEEE Transactions on Robotics and Automation 14(6):926–939. https://doi.org/10.1109/70.736776.
Benjamin M. R. 2002. Interval programming: A multi-objective optimization model for autonomous vehicle control. Doctoral dissertation Brown University;
Benjamin M. R, Leonard J. J, Curcio J. A, Newman P. M. 2006;A method for protocol-based collision avoidance between autonomous marine surface craft. Journal of Field Robotics 23(5):333–346. https://doi.org/10.1002/rob.20121.
Benjamin M. R, Newman P. M, Schmidt H, Leonard J. J. 2009. An overview of MOOS-IvP and a brief users guide to the IvP Helm autonomy software (MIT-CSAIL-TR-2009-028) Massachusetts Institute of Technology Computer Science and Artificial Intelligence Laboratory.
Benjamin M. R, Schmidt H, Newman P. M, Leonard J. J. 2010;Nested autonomy for unmanned marine vehicles with MOOS-IvP. Journal of Field Robotics 27(6):834–875. https://doi.org/10.1002/rob.20370.
Brooks R. A. 1986;A robust layered control system for a mobile robot. IEEE Journal on Robotics and Automation 2(1):14–23. https://doi.org/10.1109/JRA.1986.1087032.
Cho Y, Kim J, Kim J, Jo Y, Ryu J. 2019;Behavior-based control considering the interaction between a human operator and an autonomous surface vehicle. Journal of Ocean Engineering and Technology 33(6):620–626. https://doi.org/10.26748/KSOE.2019.093.
Elkins L, Sellers D, Monach W. R. 2010;The Autonomous Maritime Navigation (AMN) Project: Field tests, autonomous and cooperative behaviors, data fusion, sensors, and vehicles. Journal of Field Robotics 27(6):790–818. https://doi.org/10.1002/rob.20367.
Fossen T. I. 2011. Handbook of marine craft hydrodynamics and motion control John Wiley & Sons.
Gratzer G. 2002. General lattice theory Birkhauser.
Gratzer G. 2011. Lattice theory: foundation Birkhauser.
Khatib O. 1986;Real-time obstacle avoidance for manipulators and mobile robots. The International Journal of Robotics Research 5(1):90–98. https://doi.org/10.1177/027836498600500106.
Kim U. 2023. Behavior-based control for marine vehicles using task composition. Master’s thesis Korea Advanced Institute of Science and Technology (KAIST);
Kortenkamp D, Simmons R, Brugali D. 2016. Robotic systems architectures and programming. In : Siciliano B, Khatib O, eds. Springer Handbook of Robotics 2nd edth ed. p. 283–306. Springer.
Nangue Tasse G, James S, Rosman B. 2020;A Boolean task algebra for reinforcement learning. Advances in Neural Information Processing Systems 33:9497–9507.
Pirjanian P. 1999. Behavior coordination mechanisms - state-of-the-art (Tech. Rep. IRIS-99-375) University of Southern California.
Schaul T, Horgan D, Gregor K, Silver D. 2015;Universal value function approximators. Proceedings of the 32nd International Conference on Machine Learning, PMLR 37:1312–1320.
Sutton R. S, Barto A. G. 1998;Reinforcement learning: An introduction MIT Press;
Van Niekerk B, James S, Earle A, Rosman B. 2019;Composing value functions in reinforcement learning. Proceedings of the 36th International Conference on Machine Learning, PMLR 97:6401–6409.
Van der Pol E, Oliehoek F. A. 2014. Multi-objective reinforcement learning Department of Computer Science, University of Amsterdam.
Wolf M. T, Rahmani A, De La Croix J. P, Woodward G, Hook J. V, Brown D, Schaffer S, Lim C, Bailey P, Tepsuporn S, Pomerantz M, Verma V, Cain C, Cherng D. 2017;CARACaS multi-agent maritime autonomy for unmanned surface vehicles in the Swarm II harbor patrol demonstration. Proceedings of SPIE 10195, Unmanned Systems Technology XIX 1019507https://doi.org/10.1117/12.2262067.

Article information Continued

Funded by : Defense Acquisition Program Administration and the Agency for Defense Development
Award ID : 915057401
Funding : This study was supported by the Defense Acquisition Program Administration and the Agency for Defense Development under Grant 915057401.

Fig. 1

Comparison of robot control architectures: (a) Traditional deliberative architecture based on a sequential sense-plan-act pipeline, and (b) the proposed behavior-based control architecture, which processes perception data through parallel, concurrent behaviors for reactive responses

Fig. 2

Hierarchical structure of task composition. The overarching mission is logically composed of multiple independent subtasks, where each sub-task represents a specific behavior encoded as an objective function. scalar-weight tuning step required whenever the behavior set is modified. Fig. 2 presents the resulting hierarchical structure of task composition, in which the overarching mission is logically composed of multiple independent sub-tasks, each corresponding to a behavior encoded as an objective function.

Fig. 3

Behavior-based control architecture for USV. Environmental perception triggers a parallel bank of behaviors within the behavior manager. The behavior arbitrator combines their respective objective functions to generate an optimal guidance signal, which is executed by controllers driving the USV dynamics.

Fig. 4

Mechanism of the formation control behavior. The desired position for the formation is derived using an average-based method (Balch & Arkin, 1998). The position error relative to the desired formation point is decoupled into a fore-aft error regulated by speed control, and a lateral error regulated by heading control.

Fig. 5

Operator semantics over the action space: (left to right) two base tasks M1, M2, the disjunction M1∨M2 (pointwise max), the conjunction M1∧M2 (pointwise min), and the negation ¬M1 = 1 − M1. Brighter regions correspond to higher objective values; the star marker denotes the action selected by argmax over each composite map.

Fig. 6

Comparison between Task Composition (algebraic disjunction) and Linear Composition (weighted summation) in a multi-target rescue scenario. (a, b) Individual rescue objectives (MR1, MR2); (c) Disjunction preserves the multimodality of the action space (MR1∨MR2), allowing commitment to a specific target; (d) Linear Composition improperly averages the objectives (0.5 MR1 + 0.5 MR2), resulting in an invalid action straight between the targets.

Fig. 7

Action-space conjunction for simultaneous waypoint tracking and obstacle avoidance. (a) Waypoint tracking objective (Mwpt); (b) Collision avoidance objective (Mavoidance); (c) Conjunction (Mwpt ∧ Mavoidance) steers a safe evasive maneuver.

Fig. 8

Time-lapse visualization of the integrated multi-agent scenario. Four USVs successfully maintain a horizontal formation while dynamically breaking off to rescue scattered targets (green squares) and actively avoiding static obstacles (black circles), showing robust multi-objective arbitration.

Table 1

Simulation parameters.

Parameter Value
Scenario (randomized per trial)

Own-ship agents (line-abreast) 4
Formation waypoint (150, 1700) m
Static obstacles (count, x, y, radius) 10, x∈[0,400] m, y∈ [400,1600] m, r∈[30,40] m
Rescue targets (count, x, y) 5, x∈[−50,350] m, y∈[500,1500] m

Behavior parameters

Formation thresholds (dead / control) 20 / 40 m
Formation Kp (dead / control / ballistic) 0.01 / 0.05 / 0.07
Nominal cruise speed 5 m/s
Inter-vessel avoidance range 80 m

Table 2

Behavior weights used in the Linear Composition (baseline).

Behavior k Weight wk
Formation tracking 1.0
Static obstacle avoidance 3.0
Rescue 1.5

Table 3

Results of the integrated scenario (1,000 trials)

Algorithm Linear Composition Task Composition (proposed)
Position error (m) 31.63 [29.23, 34.03] 22.78 [22.31, 23.25]
Rescue rate (%) 81.74 [79.23, 84.01] 90.32 [88.33, 92.00]
Collision rate (%) 4.56 [3.43, 6.04] 0.85 [0.44, 1.64]