Automatic Multi-Sections Control Overview

• Automatic multi-sections control is defined as the coordinated management of interconnected control partitions, addressing challenges like cross-section coupling and switching transients.
• Architectural approaches include centralized multi-output control, synchronized parallel control, hierarchical coordination, and distributed sampled-data strategies tailored to domain-specific needs.
• Empirical evidence shows that tailored methods, such as curriculum-based RL and fuzzy inference, can enhance disturbance rejection, spray precision, and traffic weaving efficiency.

Searching arXiv for papers on automatic multi-section(s) control and closely related control architectures. Automatic multi-sections control denotes a class of control problems in which multiple sections, spans, lanes, controller paths, or resource partitions must be coordinated automatically rather than treated as isolated loops. In the cited literature, the term appears explicitly in agricultural spraying, and closely related formulations arise in roll-to-roll (R2R) manufacturing, multi-section continuum robots, weaving sections in mixed traffic, synchronized multi-mode control paths, humanoid robot resource management, and leader-following sampled-data multi-agent systems. Across these settings, the common technical issues are cross-section coupling, switching transients, uncertainty, and the need to reconcile global coordination with local actuation (Plessen, 15 Aug 2025, Li et al., 27 Feb 2026, Liu et al., 2024, Yan et al., 2024, Mirafzal et al., 2023, Moskovskaya et al., 12 Sep 2025, Liu et al., 2020).

1. Problem class and domain span

The literature covers several distinct meanings of “section.” In a three-section R2R line, the sections are processing spans with independent torque control but coupled web tension and roller velocity; in agricultural spraying, the sections are boom segments or individual nozzles; in a three-section continuum robot, the sections are serial cable-driven bodies; in weaving traffic, the controlled region is partitioned into cells and coordinated through section-level weights; in synchronized parallel control, the “sections” are concurrent controller paths; and in social robotics, control is partitioned across body-part agents with shared resources (Li et al., 27 Feb 2026, Plessen, 15 Aug 2025, Liu et al., 2024, Yan et al., 2024, Mirafzal et al., 2023, Moskovskaya et al., 12 Sep 2025).

Domain	Controlled sections or paths	Dominant control issue
R2R manufacturing	three processing spans / rollers	strong coupling, disturbance propagation
Agricultural spraying	48, 2, or 1 boom sections	overlap reduction, turn compensation, localization
Continuum robotics	three sections, 6-DOF	multiple IK solutions, hysteresis, cable elongation
Weaving traffic	cells and per-CAV local controllers	mandatory lane changes, dynamic topology
Multi-mode control	primary and auxiliary paths	seamless switching
Social robotics	agent forest over robot subsystems	concurrency and conflict resolution

In the R2R benchmark, Section 1 is the first processing span after the unwinder, Section 2 is the middle span and the most important for the disturbance test, and Section 3 is the last processing span before the rewinder. Because web tension in one span affects neighboring spans, the system is strongly coupled, making it a genuine multi-section MIMO control problem (Li et al., 27 Feb 2026). In the continuum-robot setting, the control challenge is likewise coupling-dominated: the robot has three sections, each with 2 DoF, and the task is to track desired tip position and orientation despite multiple IK solutions and nonlinearities such as hysteresis, cable elongation, friction, and external disturbances (Liu et al., 2024).

In agricultural spraying, automatic Section Control (ASC) is presented as the traditional high-precision approach for minimizing overlaps by switching off nozzles over already sprayed areas and dynamically adjusting nozzle flow rates when velocities of boom sections vary during turn maneuvers. The same paper also emphasizes that spraying is highly dynamic and uncertain because of machinery speed changes, boom oscillations and inertia, boom height, nozzle spacing, nozzle spray angle, crosswind and drift, clogging, PWM flow-rate effects, field geometry, path planning, and accurate localization requirements (Plessen, 15 Aug 2025). In weaving sections, the central problem is that intensive mandatory lane changes generate interactions, friction, and capacity drop, and purely local control cannot coordinate the whole section efficiently (Yan et al., 2024). In humanoid tour guidance, the problem is not only authoring behaviors but executing overlapping speech, gaze, gesture, and facial actions without blocking or abrupt transitions (Moskovskaya et al., 12 Sep 2025).

A plausible synthesis is that automatic multi-sections control is less a single algorithm than a recurring systems problem: multiple coupled control partitions must be regulated jointly under uncertainty, switching, or shared-resource constraints.

2. Architectural forms

One major architectural pattern is centralized multi-output control. The R2R paper formulates the problem as a single-agent continuous-control MDP

$$(\mathcal{S}, \mathcal{A}, \mathcal{P}, \mathcal{R}, \gamma)$$

in which a centralized policy controls all roller torques jointly. For $N=3$, the observation is a 22-dimensional state,

$$s_t = [\mathbf{T}_{norm}, \mathbf{v}_{norm}, \mathbf{T}^{ref}_{norm}, \mathbf{v}^{ref}_{norm}, \mathbf{e}_T, \mathbf{e}_v, \mathbf{u}_{t-1,norm}, \phi_t]^T \in \mathbb{R}^{7N+1},$$

and the action is

$a_t = [u_1, u_2, u_3]^T \in [-1,1]^3.$

The paper is explicit that it trains “a single policy to handle both nominal tracking and disturbance rejection,” with no separate per-section controllers (Li et al., 27 Feb 2026).

A second pattern is synchronized parallel control. “Synced Parallel Control Paths” uses a dual-pathway design with an active primary path and one or more auxiliary paths running in parallel and continuously synchronized. The controller-path outputs are written as

$X_{c1}(s)=\{w(R_1-R_1)+\bar{u}\,e(s)\}G_{c1}(s),$

$X_{c2}(s)=\{\bar{w}(R_2-R_2)+u\,e(s)\}G_{c2}(s),$

with

$e(s)=(uX_{c1}(s)+\bar{u}X_{c2}(s)),$

and total controller output

$Y_c(s)=wX_{c1}(s)+\bar{w}X_{c2}(s).$

The inactive path is not dormant; it is synchronized through an error loop so that when switching occurs there is no abrupt change in plant input (Mirafzal et al., 2023).

A third pattern is hierarchical or bi-level coordination. In weaving control, the upper level is a roadside DRL controller that observes the entire weaving section and chooses control weights, while the lower level is an MPC inside each CAV that computes acceleration and steering based on the local situation. The upper level therefore performs system-wide coordination without directly prescribing trajectories (Yan et al., 2024). In humanoid tour guidance, the execution layer is a multi-agent forest: a dispatcher ROS node initializes agents, defines their hierarchy, and forwards commands, while each agent controls a body part or subsystem, maintains its own internal state, and resolves conflicts locally (Moskovskaya et al., 12 Sep 2025).

A fourth pattern is distributed sampled-data coordination. In leader-following consensus, each follower obeys

$\dot{x}_i = A x_i + B u_i,\qquad i=1,\dots,N,$

with leader dynamics

$\dot{x}_0 = A x_0,$

and sampled-data control law

$N=3$0

This formulation is sectioned by agent index rather than physical spans, but it addresses the same coordination issue: multiple subsystems must align to a common objective under sampling constraints (Liu et al., 2020).

3. Mathematical structure of coupling and control

The mathematical hallmark of multi-sections control is explicit interdependence between neighboring or competing sections. In the three-section R2R benchmark, the span and roller dynamics are

$N=3$1

$N=3$2

These equations show the coupling explicitly: tension depends on neighboring velocities and tensions, while velocity depends on tension difference across adjacent spans. The system is discretized with Euler integration at $N=3$3 s, and control inputs include additive Gaussian noise with $N=3$4 (Li et al., 27 Feb 2026).

In agricultural spraying, coupling enters through boom-section kinematics during turns. For section $N=3$5,

$N=3$6

and the actual spray volume rate is modeled as

$N=3$7

Uniform target spray volume requires

$N=3$8

and therefore

$N=3$9

The paper also notes special cases: if $$s_t = [\mathbf{T}_{norm}, \mathbf{v}_{norm}, \mathbf{T}^{ref}_{norm}, \mathbf{v}^{ref}_{norm}, \mathbf{e}_T, \mathbf{e}_v, \mathbf{u}_{t-1,norm}, \phi_t]^T \in \mathbb{R}^{7N+1},$$0, a section may be switched off; if a section overlaps already sprayed area, it can be switched off; and in the 1-section case all nozzles share the same signal (Plessen, 15 Aug 2025).

In sampled-data multi-agent consensus, the closed-loop error dynamics for the static graph case are

$$s_t = [\mathbf{T}_{norm}, \mathbf{v}_{norm}, \mathbf{T}^{ref}_{norm}, \mathbf{v}^{ref}_{norm}, \mathbf{e}_T, \mathbf{e}_v, \mathbf{u}_{t-1,norm}, \phi_t]^T \in \mathbb{R}^{7N+1},$$1

where $$s_t = [\mathbf{T}_{norm}, \mathbf{v}_{norm}, \mathbf{T}^{ref}_{norm}, \mathbf{v}^{ref}_{norm}, \mathbf{e}_T, \mathbf{e}_v, \mathbf{u}_{t-1,norm}, \phi_t]^T \in \mathbb{R}^{7N+1},$$2 and $$s_t = [\mathbf{T}_{norm}, \mathbf{v}_{norm}, \mathbf{T}^{ref}_{norm}, \mathbf{v}^{ref}_{norm}, \mathbf{e}_T, \mathbf{e}_v, \mathbf{u}_{t-1,norm}, \phi_t]^T \in \mathbb{R}^{7N+1},$$3. The design uses

$$s_t = [\mathbf{T}_{norm}, \mathbf{v}_{norm}, \mathbf{T}^{ref}_{norm}, \mathbf{v}^{ref}_{norm}, \mathbf{e}_T, \mathbf{e}_v, \mathbf{u}_{t-1,norm}, \phi_t]^T \in \mathbb{R}^{7N+1},$$4

with $$s_t = [\mathbf{T}_{norm}, \mathbf{v}_{norm}, \mathbf{T}^{ref}_{norm}, \mathbf{v}^{ref}_{norm}, \mathbf{e}_T, \mathbf{e}_v, \mathbf{u}_{t-1,norm}, \phi_t]^T \in \mathbb{R}^{7N+1},$$5 solving

$$s_t = [\mathbf{T}_{norm}, \mathbf{v}_{norm}, \mathbf{T}^{ref}_{norm}, \mathbf{v}^{ref}_{norm}, \mathbf{e}_T, \mathbf{e}_v, \mathbf{u}_{t-1,norm}, \phi_t]^T \in \mathbb{R}^{7N+1},$$6

The key convergence condition is

$$s_t = [\mathbf{T}_{norm}, \mathbf{v}_{norm}, \mathbf{T}^{ref}_{norm}, \mathbf{v}^{ref}_{norm}, \mathbf{e}_T, \mathbf{e}_v, \mathbf{u}_{t-1,norm}, \phi_t]^T \in \mathbb{R}^{7N+1},$$7

This formalism makes the effect of sampling-induced mismatch explicit (Liu et al., 2020).

In weaving control, the upper-level controller outputs cell-dependent weights

$$s_t = [\mathbf{T}_{norm}, \mathbf{v}_{norm}, \mathbf{T}^{ref}_{norm}, \mathbf{v}^{ref}_{norm}, \mathbf{e}_T, \mathbf{e}_v, \mathbf{u}_{t-1,norm}, \phi_t]^T \in \mathbb{R}^{7N+1},$$8

which are used by the lower-level MPC objective

$$s_t = [\mathbf{T}_{norm}, \mathbf{v}_{norm}, \mathbf{T}^{ref}_{norm}, \mathbf{v}^{ref}_{norm}, \mathbf{e}_T, \mathbf{e}_v, \mathbf{u}_{t-1,norm}, \phi_t]^T \in \mathbb{R}^{7N+1},$$9

Safety is handled explicitly in MPC constraints rather than only by objective weighting (Yan et al., 2024).

4. Methodological families

The cited work spans several methodological families rather than a single dominant paradigm. In R2R manufacturing, the central method is curriculum-based Soft Actor-Critic. The SAC objective is

$a_t = [u_1, u_2, u_3]^T \in [-1,1]^3.$0

with entropy term

$a_t = [u_1, u_2, u_3]^T \in [-1,1]^3.$1

The implementation uses one stochastic actor $a_t = [u_1, u_2, u_3]^T \in [-1,1]^3.$2, two critics $a_t = [u_1, u_2, u_3]^T \in [-1,1]^3.$3, $a_t = [u_1, u_2, u_3]^T \in [-1,1]^3.$4, soft target networks, automatic temperature tuning, Polyak averaging with $a_t = [u_1, u_2, u_3]^T \in [-1,1]^3.$5, and action smoothing

$a_t = [u_1, u_2, u_3]^T \in [-1,1]^3.$6

with $a_t = [u_1, u_2, u_3]^T \in [-1,1]^3.$7. Its distinctive contribution is a three-phase curriculum: Phase 1 uses tension range 27–33 N, velocity range 0.0095–0.0105 m/s, and step change probability 10%; Phase 2 uses 25–35 N, 0.009–0.011 m/s, and 20%; Phase 3 uses 20–40 N, 0.008–0.012 m/s, and 30% (Li et al., 27 Feb 2026).

In continuum robotics, the methodological choice is almost the opposite: a model-less Mamdani fuzzy inference system with 11 input variables and 10 output variables. Its three stages are fuzzification, inference, and defuzzification. To keep the rule base small, each input uses only two membership functions, Z-shaped membership function for negative error and S-shaped membership function for positive error. The controller directly maps task-space error to cable-length changes and includes a built-in shape reconstruction algorithm that uses only tip position and orientation feedback together with the sign of selected middle-section cable displacements (Liu et al., 2024).

In mixed traffic weaving, the method is bi-level. The upper level uses PPO with an actor–critic architecture and CNN layers because the state is spatially structured; the lower level uses MPC with explicit safety constraints, a binary lane-change variable, and an HV trajectory predictor based on EvolveGCN, GIN, CBAM, and CNN. The predictor is trained with an $a_t = [u_1, u_2, u_3]^T \in [-1,1]^3.$8 loss, and the upper level optimizes speed, flow, and missed-weaving penalties via an RL reward (Yan et al., 2024).

In agricultural spraying, the methodological contrast is between fine-grained sensor-based ASC and a simpler predictive spray switching alternative. The alternative method precomputes where spray should be on or off based on path planning and transition geometry rather than reacting only to already sprayed cells in an occupancy grid. The paper also replaces occupancy-grid overlap detection with a recursive polygon representation of sprayed area, retaining only the last 10 sampling spaces for efficiency (Plessen, 15 Aug 2025).

This diversity of methods suggests that multi-sections control is organized more by structural constraints than by allegiance to a single control paradigm. Learning-based, fuzzy, predictive, hierarchical, and distributed sampled-data approaches all appear when the section-interaction pattern demands them.

5. Performance and empirical evidence

The most quantitatively detailed R2R results concern both nominal tracking and large disturbances. On the nominal $a_t = [u_1, u_2, u_3]^T \in [-1,1]^3.$9 N constant-tension task across all sections, best SAC performance over 10 episodes is an episode reward of 4.9809, tension MAE of 0.0036 N, tension RMSE of 0.0041 N, velocity MAE of 0.000130 m/s, and velocity RMSE of 0.000140 m/s. The corresponding MPC and LQR rewards are 4.6883 and 4.7129. In the disturbed case, Section 2 experiences a $X_{c1}(s)=\{w(R_1-R_1)+\bar{u}\,e(s)\}G_{c1}(s),$0 step at $X_{c1}(s)=\{w(R_1-R_1)+\bar{u}\,e(s)\}G_{c1}(s),$1 s while Sections 1 and 3 remain at 30 N. For Section 2, SAC MAE is 0.336 N, compared with 0.510 N for MPC and 0.347 N for LQR; rise time is 0.12 s, settling time is 0.31 s, and overshoot is 2.3%. For coupling rejection in Sections 1 and 3, SAC maximum deviation is 0.82 N versus 1.54 N for MPC and 1.18 N for LQR (Li et al., 27 Feb 2026).

The continuum-robot experiments are physical rather than purely simulated. On a three-section, six-degree-of-freedom continuum robot of total length 165 mm and outer diameter 12.7 mm, with VICON feedback and a 4 mm silicone rod inserted as an internal payload simulator, the proposed FLC achieves circle-path RMSE of 0.28 mm and dual-loop RMSE of 0.54 mm. These correspond to about 0.17%–0.32% of robot length. Under an unexpected 100 g weight during tracking, the FLC system recovers quickly and remains stable, whereas the PCC-based controller produces large oscillations and diverges (Liu et al., 2024).

The weaving-section case study is calibrated to a real site near Basel, Switzerland, with length 535 m, two main lanes plus one auxiliary lane, evening-peak demand from 16:00–17:00, and time step 0.2 s. Across all CAV penetration rates, Bi-Level-G achieves the highest space-mean speed and exit flow. At 20% CAV penetration, it reaches 60.35 km/h and 1743 veh/(lane·h); at 100% penetration, it reaches 85.80 km/h and 1877 veh/(lane·h). Relative to HDM, space-mean-speed improvements reach 54.73%, and exit flow improves by up to 12.86% (Yan et al., 2024).

In agricultural spraying, the preferred path-planning method $X_{c1}(s)=\{w(R_1-R_1)+\bar{u}\,e(s)\}G_{c1}(s),$2 always outperforms Boustrophedon $X_{c1}(s)=\{w(R_1-R_1)+\bar{u}\,e(s)\}G_{c1}(s),$3 in pathlength, with savings ranging from $X_{c1}(s)=\{w(R_1-R_1)+\bar{u}\,e(s)\}G_{c1}(s),$4 to $X_{c1}(s)=\{w(R_1-R_1)+\bar{u}\,e(s)\}G_{c1}(s),$5. For spray volume, the 48-section setup comes closest to the ideal reference. In field 1, the 48-section case is about 2.1% above reference, the 2-section case about 21.0% above, and the 1-section case about 26.9% above. The same trend appears across all 10 fields: 48 sections yield the smallest excess spray volume, 2 sections are intermediate, and 1 section has the largest excess spray volume (Plessen, 15 Aug 2025).

The synchronized-parallel control paper offers simulation evidence rather than theorem-heavy guarantees. In Example (1), synchronized loops yield stable and coordinated output, whereas disabling synchronization by setting $X_{c1}(s)=\{w(R_1-R_1)+\bar{u}\,e(s)\}G_{c1}(s),$6 produces irregular behavior. In Example (2), the synchronized case again yields coordinated and stable responses, while the unsynchronized case is more oscillatory (Mirafzal et al., 2023).

6. Tradeoffs, limitations, and recurrent controversies

A recurrent controversy is whether finer sectional granularity is always the preferred engineering choice. The agricultural spraying study answers negatively under nominal conditions. It states that 48-section ASC gives the best spray-volume precision and lowest overlap, but still recommends the alternative predictive switching method with 1 or 2 sections as a practical solution because it minimizes pathlength, offers intermediate overlap, can be done sensor-free, is low-cost, and is suitable for manual driving. This recommendation is explicitly a tradeoff, not a claim that the simpler method beats ASC in pure technical precision (Plessen, 15 Aug 2025).

A second recurring misconception is that multi-sections control necessarily implies one controller per section. The R2R study rejects that interpretation by using one centralized SAC policy for all three rollers and by emphasizing that the same learned policy handles nominal and disturbed conditions without scenario-specific retuning. The paper also notes a limitation: although the method uses process noise and operational diversity via curriculum, it does not randomize physical parameters like $X_{c1}(s)=\{w(R_1-R_1)+\bar{u}\,e(s)\}G_{c1}(s),$7 or $X_{c1}(s)=\{w(R_1-R_1)+\bar{u}\,e(s)\}G_{c1}(s),$8; full parameter-uncertainty domain randomization is identified as future work (Li et al., 27 Feb 2026).

A third issue concerns switching. In multi-mode systems, the central problem is not controlling each mode individually but switching between modes without transients, loss of regulation, or instability. “Synced Parallel Control Paths” addresses this by keeping auxiliary paths synchronized continuously rather than idle, and it notes an implementation detail: typically $X_{c1}(s)=\{w(R_1-R_1)+\bar{u}\,e(s)\}G_{c1}(s),$9 and $X_{c2}(s)=\{\bar{w}(R_2-R_2)+u\,e(s)\}G_{c2}(s),$0 are PI controllers, while non-PI cases require an integrator in the error path to eliminate steady-state mismatch (Mirafzal et al., 2023). The sampled-data consensus paper raises a related but more formal caution: a previously claimed equivalence between two control laws holds only when sampling is synchronous for all subsystems, that is, $X_{c2}(s)=\{\bar{w}(R_2-R_2)+u\,e(s)\}G_{c2}(s),$1 for all $X_{c2}(s)=\{\bar{w}(R_2-R_2)+u\,e(s)\}G_{c2}(s),$2. The same paper extends the convergence guarantee from periodic sampling to bounded aperiodic sampling, but only under the explicit upper-bound condition on $X_{c2}(s)=\{\bar{w}(R_2-R_2)+u\,e(s)\}G_{c2}(s),$3 and a positive lower bound $X_{c2}(s)=\{\bar{w}(R_2-R_2)+u\,e(s)\}G_{c2}(s),$4 (Liu et al., 2020).

A fourth issue is the allocation of intelligence between content generation and execution. In humanoid tour guidance, the LLM generates the narrative and action tags, but the robot control layer owns execution logic. The paper explicitly identifies hallucinations, synchronization issues, and returning to default behavior as remaining problems in LLM-generated scenarios, and assigns these to the multi-agent execution system rather than to the LLM itself (Moskovskaya et al., 12 Sep 2025).

Taken together, these results indicate that automatic multi-sections control is best understood as a coordination problem under coupling, switching, and uncertainty. The literature does not support a single universal solution. Instead, it shows a spectrum of viable designs: centralized RL for tightly coupled spans, fuzzy model-less control for nonlinear continuum sections, bi-level DRL–MPC for bottlenecks with mixed autonomy, synchronized parallel paths for bumpless mode transfer, distributed sampled-data designs for leader-following coordination, and coarse but practical predictive switching when sensing and implementation cost dominate.