Distributed Manipulator Systems (DMS)

• Distributed Manipulator Systems (DMS) are robotic architectures that integrate distributed actuation, sensing, and control across cooperating subsystems for coordinated object manipulation.
• They leverage diverse methodologies including graph-based formations, potential fields, and energy-flow models to manage coupling constraints and ensure synchronized movement.
• Applications span cooperative transport, underwater manipulation, and modular reconfigurable platforms, highlighting challenges in sensing integration, planning, and system scalability.

Distributed Manipulator Systems (DMS) are robotic manipulation architectures in which actuation, sensing, or both are distributed across multiple cooperating subsystems rather than concentrated in a single fixed-base arm. In the recent robotics literature, the term covers several closely related embodiments: teams of mobile or floating manipulators rigidly grasping a common object, networks of manipulators coordinating end-effector formations, actuator arrays that manipulate through a shared surface, and manipulators whose actuation is itself distributed across links or reconfigurable modules (Sewlia et al., 16 Dec 2025, Liu et al., 2024, Dacre et al., 23 Feb 2026, Nishio et al., 2024). Across these variants, the unifying technical issue is coupling: the motion, wrench, or geometric state of one subsystem constrains the feasible behavior of the others.

1. Scope, terminology, and system classes

In robotics usage, DMS refers to manipulation systems with distributed physical embodiment, distributed sensing, distributed control, or some combination thereof. This should be distinguished from other uses of the acronym: in distributed computing it can denote a Demand Migration System, in power systems a Distribution Management System, and in manufacturing a Distributed Manufacturing Simulation (Ji et al., 2010, Kundacina et al., 2023, Saad et al., 2019). Within robotics, however, the literature surveyed here uses DMS to describe cooperative manipulation teams, distributed actuator arrays, and distributed-actuation manipulator bodies.

A recurring ambiguity in the robotics literature is that “distributed” does not always mean the same thing. Some works are distributed only in the physical sense: multiple manipulators or actuator tiles exist, but planning and coordination are centralized. Other works are distributed in the stronger control-theoretic sense: each subsystem computes locally from neighbor information, local sensing, or object-mediated interaction, without centralized optimization or full-state broadcast (Sewlia et al., 16 Dec 2025, Liu et al., 2024).

DMS class	Coupling mechanism	Representative papers
Cooperative mobile or floating manipulators	Rigid grasp of a common object	(Sewlia et al., 16 Dec 2025, Heshmati-Alamdari et al., 2019, Heshmati-Alamdari et al., 2019)
Networked manipulator formations	Graph-based relative-pose or distance constraints	(Ge et al., 2016, Wu et al., 2021)
Distributed actuator arrays	Compliant or interconnected manipulation surfaces	(Dacre et al., 23 Feb 2026, Dacre et al., 17 Sep 2025)
Distributed-actuation or reconfigurable bodies	Thrust or structure distributed across the manipulator	(Nishio et al., 2024, Nguyen et al., 31 Oct 2025)
Distributed sensing layers for manipulation	Local tactile or inductive sensing with local inference	(Dacre et al., 3 Feb 2025, He et al., 2023)

The breadth of these embodiments matters because DMS is not confined to one morphology. In one extreme, the system is a closed-chain team of mobile manipulators carrying a payload through narrow passages (Sewlia et al., 16 Dec 2025). In another, it is a surface-scale array of 3-DoF tiles that manipulates by shaping both rigid end-effectors and the compliant material between them (Dacre et al., 23 Feb 2026). The common problem is coordinated manipulation under coupling constraints that are distributed in space.

2. Mathematical structure of coupling

A large part of DMS theory begins from manipulator or mobile-manipulator dynamics of Euler–Lagrange form. For a mobile manipulator $i$, one formulation uses generalized coordinates $$q_i = \begin{bmatrix} b_i & \theta_i \end{bmatrix}$$ and dynamics

$$M_i(q_i)\ddot{q}_i + C_i(q_i,\dot{q}_i)\dot{q}_i + g_i(q_i) = \tau_i,$$

with $b_i$ the planar base position and $\theta_i$ the arm joints (Sewlia et al., 16 Dec 2025). Closely related formulations appear for fixed-base manipulator networks and mobile-manipulator teams, where each subsystem is modeled as a second-order nonlinear mechanical system and the coordination objective is imposed either in joint space or in task space (Ge et al., 2016, Wu et al., 2021).

When multiple manipulators rigidly grasp a common object, coupling becomes explicit through grasp compatibility and shared payload dynamics. One representative model assumes constant relative pose between each end-effector frame and the object frame, yielding the velocity compatibility constraint

$J(q)\dot{q} = G^\top v_o,$

and wrench relation

$\lambda_o = G\lambda.$

This leads to coupled object–robot dynamics of the form

$$\tilde{M}\dot{v}_o + \tilde{C}v_o + \tilde{g} = G^\top u,$$

which makes the payload the locus of coordination for the entire team (Sewlia et al., 16 Dec 2025). Underwater cooperative-manipulation formulations express the same structure through $v_i = J_{iO}v_O$, $\lambda_O = G^\top\lambda$, and local decompositions of object dynamics using coefficients $$q_i = \begin{bmatrix} b_i & \theta_i \end{bmatrix}$$0, $$q_i = \begin{bmatrix} b_i & \theta_i \end{bmatrix}$$1, so that each agent carries an assigned share of inertia, Coriolis, damping, and gravity terms (Heshmati-Alamdari et al., 2019).

A second major coupling mechanism is graph-theoretic rather than grasp-theoretic. In distributed formation tracking, the communication graph $$q_i = \begin{bmatrix} b_i & \theta_i \end{bmatrix}$$2, its adjacency matrix, degree matrix, and Laplacian encode which manipulators exchange information. One line of work uses $$q_i = \begin{bmatrix} b_i & \theta_i \end{bmatrix}$$3 to characterize leader reachability and finite-time formation tracking (Ge et al., 2016). Another uses the directed-graph Laplacian $$q_i = \begin{bmatrix} b_i & \theta_i \end{bmatrix}$$4 to coordinate force-induced correction states under bounded asymmetric delays (Liu et al., 2024). In both cases, the graph replaces the payload as the explicit medium of coupling.

Task-space DMS formulations also frequently use potential-based geometric coupling. For fixed-base manipulator formations, edge variables $$q_i = \begin{bmatrix} b_i & \theta_i \end{bmatrix}$$5 and edge errors $$q_i = \begin{bmatrix} b_i & \theta_i \end{bmatrix}$$6 induce a potential $$q_i = \begin{bmatrix} b_i & \theta_i \end{bmatrix}$$7, whose gradient $$q_i = \begin{bmatrix} b_i & \theta_i \end{bmatrix}$$8 acts as a virtual spring force at end-effector $$q_i = \begin{bmatrix} b_i & \theta_i \end{bmatrix}$$9. Mapping this through a Jacobian transpose produces joint torques that realize task-space formation regulation (Wu et al., 2021). This formulation is notable because it embeds multi-agent formation control directly into manipulator dynamics rather than assuming point-mass agents.

At a more structural level, vehicle-manipulator systems admit a port-Hamiltonian formulation in which the state $$M_i(q_i)\ddot{q}_i + C_i(q_i,\dot{q}_i)\dot{q}_i + g_i(q_i) = \tau_i,$$0 evolves according to

$$M_i(q_i)\ddot{q}_i + C_i(q_i,\dot{q}_i)\dot{q}_i + g_i(q_i) = \tau_i,$$1

with skew interconnection matrix $$M_i(q_i)\ddot{q}_i + C_i(q_i,\dot{q}_i)\dot{q}_i + g_i(q_i) = \tau_i,$$2, Hamiltonian $$M_i(q_i)\ddot{q}_i + C_i(q_i,\dot{q}_i)\dot{q}_i + g_i(q_i) = \tau_i,$$3, and power balance

$$M_i(q_i)\ddot{q}_i + C_i(q_i,\dot{q}_i)\dot{q}_i + g_i(q_i) = \tau_i,$$4

This makes energy flow explicit and separates the Lie–Poisson structure of the base from the canonical symplectic structure of the manipulator (Rashad, 25 Feb 2026). Although derived for single vehicle-manipulator systems, the formalism is directly relevant to DMS because passivity-preserving interconnection and modular power ports are natural ingredients for multi-agent manipulation.

3. Coordination and control architectures

The control literature on DMS spans distributed formation control, internal-wrench regulation, decentralized impedance control, predictive cooperative control, and centralized hierarchical planning. These are not interchangeable: each addresses a different notion of coordination.

For networked manipulator formations, one important line of work targets global finite-time convergence. A representative controller combines inverse dynamics with distributed sliding-mode estimation so that multiple $$M_i(q_i)\ddot{q}_i + C_i(q_i,\dot{q}_i)\dot{q}_i + g_i(q_i) = \tau_i,$$5-DOF manipulators achieve a prescribed time-varying formation while the formation centroid tracks a dynamic leader. Under $$M_i(q_i)\ddot{q}_i + C_i(q_i,\dot{q}_i)\dot{q}_i + g_i(q_i) = \tau_i,$$6 and $$M_i(q_i)\ddot{q}_i + C_i(q_i,\dot{q}_i)\dot{q}_i + g_i(q_i) = \tau_i,$$7, leader reachability is necessary and sufficient for the fixed-graph problem (Ge et al., 2016). A separate task-space formulation uses virtual springs between end-effectors and joint-space damping, then extends the nominal controller with an internal-model term and adaptive Jacobian estimation to handle uncertain gravity and uncertain kinematics (Wu et al., 2021). These works define DMS primarily as a graph-coupled network of manipulators.

A different control objective is regulation of unnecessary internal wrench during cooperative transport. One recent distributed law operates at the joint-velocity level, uses only local force/torque sensing and delayed neighbor information, and avoids torque control and full dynamic modeling. Its closed-loop error dynamics take the form

$$M_i(q_i)\ddot{q}_i + C_i(q_i,\dot{q}_i)\dot{q}_i + g_i(q_i) = \tau_i,$$8

with bounded asymmetric, non-uniform, time-varying delays. Under quasi-static manipulation, rigid grasp, and suitable gain and stiffness conditions, the method guarantees asymptotic convergence of correction errors and interaction wrench to the desired values (Liu et al., 2024). Here the distinguishing DMS theme is not object trajectory tracking but suppression of internal force buildup due to asynchrony and disturbances.

Underwater cooperative manipulation introduces a distinct set of constraints: bandwidth-poor communication, hydrodynamics, and cluttered 3D workspaces. One decentralized impedance framework uses a leader–follower architecture in which only the leader knows the goal and obstacle map, while followers estimate the desired object trajectory from local object motion by a prescribed-performance law

$$M_i(q_i)\ddot{q}_i + C_i(q_i,\dot{q}_i)\dot{q}_i + g_i(q_i) = \tau_i,$$9

Each robot then applies an adaptive impedance controller based on local force/torque measurements, with no explicit online data exchange among robots (Heshmati-Alamdari et al., 2019). A related but optimization-based underwater approach uses local nonlinear model predictive control for each UVMS, load-sharing coefficients $b_i$0, and common object references derived from a navigation function, again without explicit online exchange of states or controls (Heshmati-Alamdari et al., 2019). In both cases, DMS coordination is achieved through rigid grasp, object-mediated inference, and local models rather than centralized computation.

Not all effective DMS architectures are distributed in this strict sense. A notable counterexample is a multi-rate planning-and-control framework for transporting a common payload through obstacle-cluttered, highly constrained environments under STL task specifications. That system decomposes the problem into offline object-path generation, offline base-footprint planning, online constrained inverse kinematics, and continuous-time joint-space PD tracking. It yields coordinated reconfiguration of a three-mobile-manipulator team, but the planning and IK layers are centralized and hierarchical rather than decentralized (Sewlia et al., 16 Dec 2025). This distinction is consequential: many physically distributed systems remain centrally coordinated at the algorithmic level.

4. Distributed sensing, estimation, and planning

DMS performance depends on sensing architecture as much as on control law. Several papers treat centralized sensing as a bottleneck and shift estimation toward local, contact-rich, or lattice-based inference.

One explicit attempt to decentralize perception uses a soft inductive sensor array together with Neural Cellular Automata (NCA). In that formulation, each sensor cell maintains a local state $b_i$1 and updates it from immediate neighbors according to

$b_i$2

so that a global property—the object geometric center—emerges from local computation and local communication only. Hardware experiments on a $b_i$3 sensor board reported average localization errors of $b_i$4 mm and $b_i$5 mm, described in the abstract as $b_i$6 times the inter sensor distance, while synthetic tests showed size generalization from $b_i$7 to $b_i$8 grids and robustness to substantial sensor faults and noise (Dacre et al., 3 Feb 2025). This directly addresses a common DMS mismatch: distributed actuation paired with centralized sensing.

A complementary approach appears in cooperative mobile lifting. Tactile Mobile Manipulators combine a DOTS mobile robot with a TacTip tactile sensor mounted on the lifting platform, so each node can estimate local contact depth and contact angle after touching a box. In cooperative lifting with two robots, the paper reports an average success rate of $b_i$9 for the tactile system versus $\theta_i$0 for the vision-only method, while its own table averages the visual baseline to $\theta_i$1. The advantage is largest when the payload mass is non-uniformly distributed (He et al., 2023). The deeper technical point is that local contact-state matching can matter more than purely global pre-contact pose alignment.

Estimation also enters DMS control loops in less overtly sensory ways. The underwater leader–follower impedance architecture makes followers estimate the leader-induced desired object trajectory while satisfying time-varying performance envelopes (Heshmati-Alamdari et al., 2019). Formation-tracking controllers likewise reconstruct unavailable leader acceleration through distributed sliding-mode estimators before applying inverse-dynamics control (Ge et al., 2016). In both cases, estimation is not an auxiliary module but part of the distributed coordination mechanism.

Planning remains a second, orthogonal bottleneck. In constrained cooperative transport, object-level task logic and full-team kinematic feasibility do not align automatically. One hierarchical solution first generates an STL-satisfying sequence of object waypoints, then smooths the trajectory by cubic spline interpolation, Gaussian smoothing, and cubic Hermite reconstruction; next it plans collision-free base footprints with pairwise formation constraints and super-ellipse obstacle approximations; finally it solves a global constrained IK problem that enforces rigid grasp and collision penalties while allowing slight deviation from the offline base plan (Sewlia et al., 16 Dec 2025). That work is especially informative because it shows how much planning machinery is required once DMS operation enters narrow passages and clutter, but it also highlights a limitation: only object position, not full 6D object pose, is explicitly planned and tracked.

5. Embodiments and application domains

The current DMS literature spans cooperative transport teams, actuator arrays, aerial and underwater manipulators, and self-reconfigurable modular platforms. The embodiments differ sharply, but the same coordination questions recur.

Array-based DMS challenge the assumption that reliable manipulation requires dense actuator pitch. One design uses modular 3-DoF tiles arranged in an $\theta_i$2 grid and interconnected by a compliant surface layer, so objects can be manipulated not only on rigid tile end-effectors but also in the membrane spanning adjacent tiles. Feasibility is characterized through distances $\theta_i$3 and $\theta_i$4 between neighboring tile corners, with constraints $\theta_i$5 and $\theta_i$6; local region-based control then uses Dijkstra planning over tile, inter-tile, and central regions. A $\theta_i$7 prototype demonstrates translation of a cube, puck, sphere, and small rounded tetrahedron, and the authors report that their earlier interconnected-surface principle increases manipulable area by $\theta_i$8 without increasing actuator count (Dacre et al., 23 Feb 2026). A related linear-array study formulates the same interconnection geometry for 3-tile origami-inspired arrays, derives motion primitives for rolling and sliding objects, and again reports approximately $\theta_i$9 increase in traversable span without additional actuators (Dacre et al., 17 Sep 2025). These works broaden DMS beyond multi-arm grasping toward programmable manipulation surfaces.

Distributed actuation can also occur within a single robot body. A root-perching rotor-distributed manipulator distributes thrust across arm links rather than concentrating it at the aerial base, then stabilizes the root by perching on a surface, especially a ceiling. The paper argues that distributing thrust over links reduces joint torque demand, increases feasible end-effector wrench, and enlarges practical reachability relative to rotor-concentrated manipulators, while the aerial support polygon gives a geometric condition for stable roll and pitch control (Nishio et al., 2024). In a different direction, the port-Hamiltonian treatment of vehicle-manipulator systems treats aerial, underwater, space, and omnidirectional mobile manipulators as a unified class with a moving base and serial arm, providing a coordinate-free energetic model that is readily extensible to multi-arm settings (Rashad, 25 Feb 2026). Both contributions are relevant to DMS because they foreground distributed embodiment rather than only distributed algorithms.

Self-reconfiguration offers yet another route to simplifying DMS coordination. MobiDock begins from two independent mobile manipulators, then uses vision-based autonomous docking and a threaded screw-lock mechanism to create a single mobile bimanual platform. In the reported experiments, the docked system achieves RMS acceleration $J(q)\dot{q} = G^\top v_o,$0 versus $J(q)\dot{q} = G^\top v_o,$1, jerk $J(q)\dot{q} = G^\top v_o,$2 versus $J(q)\dot{q} = G^\top v_o,$3, angular standard deviation $J(q)\dot{q} = G^\top v_o,$4 versus $J(q)\dot{q} = G^\top v_o,$5, and average task time $J(q)\dot{q} = G^\top v_o,$6 s versus $J(q)\dot{q} = G^\top v_o,$7 s when compared with two independently cooperating robots (Nguyen et al., 31 Oct 2025). This is a distinct DMS strategy: instead of improving distributed control, it mechanically removes part of the coordination problem.

Cooperative mobile-manipulator transport remains the most classical application domain. Here the shared payload induces closed-chain kinematic constraints, common object dynamics, and collision-sensitive formation changes. The narrow-passage trajectory-tracking study with three Panda mobile manipulators, the distributed internal-wrench controller for teams of omnidirectional mobile manipulators, and the underwater UVMS cooperative-transport works all belong to this family, despite their different control assumptions (Sewlia et al., 16 Dec 2025, Liu et al., 2024, Heshmati-Alamdari et al., 2019). What changes across domains is not the existence of coupling but the dominant challenge: clutter and hybrid planning on land, communication scarcity underwater, or actuation allocation in aerial systems.

6. Evaluation practice, limitations, and open directions

The empirical maturity of DMS remains uneven. Several influential works provide strong simulations but limited quantitative benchmarking. The constrained-environment trajectory-tracking study demonstrates visually convincing three-robot transport through square, triangular, and trapezoidal passages, yet explicitly does not report tables of RMS tracking error, computation time per IK cycle, STL robustness margins, or ablations of the footprint planner and collision penalty (Sewlia et al., 16 Dec 2025). The finite-time formation-tracking and distributed end-effector formation papers are likewise simulation-only (Ge et al., 2016, Wu et al., 2021). Even when real-time simulation is used in ROS/UwSim for underwater cooperative transport, the validation remains simulated rather than experimental (Heshmati-Alamdari et al., 2019).

Strong assumptions also recur. Rigid grasp is pervasive in cooperative transport models (Sewlia et al., 16 Dec 2025, Heshmati-Alamdari et al., 2019, Heshmati-Alamdari et al., 2019). Quasi-static operation is central to distributed internal-wrench regulation by force sensing, which leaves impact forces and aggressive base transients outside the theory (Liu et al., 2024). Some planning frameworks mention joint-limit and singularity avoidance conceptually but do not fully specify them mathematically (Sewlia et al., 16 Dec 2025). Array-based systems often treat connective surfaces as effectively inextensible, even while acknowledging real stretch, hysteresis, snap-through, or frictional sticking (Dacre et al., 23 Feb 2026, Dacre et al., 17 Sep 2025). Self-reconfigurable systems can improve stability, but current demonstrations remain small-scale and task-specific, such as two-robot docking followed by teleoperated trash collection (Nguyen et al., 31 Oct 2025).

A second fault line is the distinction between physical distribution and computational distribution. Some systems are clearly decentralized in execution: local-information internal-wrench control under time-varying delays, leader–follower underwater impedance control with no explicit online communication, and graph-based finite-time formation control all fit that description (Liu et al., 2024, Heshmati-Alamdari et al., 2019, Ge et al., 2016). Others are physically distributed but centrally coordinated, as in hierarchical object planning plus global IK (Sewlia et al., 16 Dec 2025). Still others reduce distributedness by reconfiguring morphology, as in docking two robots into one platform (Nguyen et al., 31 Oct 2025). This distinction is not merely terminological; it determines scalability, fault modes, sensing requirements, and proof techniques.

Scale remains underexplored. The NCA-based sensing paper is unusual in demonstrating synthetic generalization to $J(q)\dot{q} = G^\top v_o,$8 agents, but its hardware remains a $J(q)\dot{q} = G^\top v_o,$9 board (Dacre et al., 3 Feb 2025). Low-density manipulation surfaces are compelling proofs of principle, yet current prototypes are $\lambda_o = G\lambda.$0 or 3-tile systems (Dacre et al., 23 Feb 2026, Dacre et al., 17 Sep 2025). TacMMs validate two-robot tactile lifting, not larger cooperative hands of mobile nodes (He et al., 2023). MobiDock validates two-module reconfiguration only (Nguyen et al., 31 Oct 2025). The literature therefore contains many scalability arguments and relatively few large-scale demonstrations.

The cited papers identify several open directions. Underwater cooperative-control works point toward richer dynamic models, better handling of impacts and accelerations, and broader validation beyond simulation (Heshmati-Alamdari et al., 2019, Heshmati-Alamdari et al., 2019). Low-density array papers identify 2D expansion, larger objects, and closed-loop transport as immediate next steps (Dacre et al., 23 Feb 2026, Dacre et al., 17 Sep 2025). MobiDock proposes autonomous task execution beyond teleoperation and extension beyond two modules (Nguyen et al., 31 Oct 2025). A plausible implication is that future DMS architectures will increasingly co-design decentralized sensing, morphology, and control rather than treating them as separable layers: local tactile or inductive inference can provide distributed state estimates, passivity-structured models can support modular interconnection, and reconfigurable or compliant embodiments can shift burden away from purely algorithmic coordination (Dacre et al., 3 Feb 2025, Rashad, 25 Feb 2026).