---
title: Multi-Arm Manipulation
url: https://www.emergentmind.com/topics/multi-arm-manipulation
type: topic
---

# Multi-Arm Manipulation

Multi-arm manipulation is the coordinated control, planning, perception, and learning of multiple robotic manipulators that jointly manipulate objects, environments, or one another’s task context. It encompasses rigid closed-chain manipulation, handovers and regrasps, articulated-object control, multi-arm task-and-motion planning, teleoperation, imitation learning, vision-language-action control, active perception, and robot design for locomotion-manipulation systems. Its defining difficulty is that multiple arms create coupled kinematic, collision, contact, temporal, and informational constraints whose feasible solutions generally occupy a substantially smaller space than the Cartesian product of independently valid single-arm behaviors.

## 1. Kinematic and mechanical foundations

When multiple arms grasp a common rigid object, the arms and object form a closed kinematic chain. If arm $i$ has configuration $\mathbf q_i$ and the object pose is $T_{\mathrm{obj}}\in SE(3)$, a composite configuration may be written as

$$
\mathbf c=(\mathbf q_1,\ldots,\mathbf q_k,T_{\mathrm{obj}}).
$$

For a prescribed object-relative grasp transform $G_i$, rigid grasping requires

$$
T_{\mathrm{ee},i}=T_{\mathrm{obj}}G_i.
$$

The feasible closed-chain set is consequently a lower-dimensional constraint manifold,

$$
\mathcal C_{\mathrm{cc}}
=
\left\{
\mathbf c\mid \mathbf F_G(\mathbf c)=0
\right\},
$$

intersected with collision-free configurations and joint-limit constraints. Direct random sampling in the ambient composite configuration space almost surely generates configurations that violate closure. This distinguishes multi-arm manipulation from merely increasing the dimensionality of single-arm motion planning. The formulation and its implications for global connectivity are developed in “Closed-Chain Manipulation of Large Objects by Multi-Arm Robotic Systems” [1612.01650].

Joint limits can disconnect the closed-chain manifold even when two configurations correspond to identical or geometrically compatible object poses. A six-axis manipulator may have up to 16 inverse-kinematic solutions for one end-effector pose, and these solutions may belong to different joint-limited components. One configuration may place an arm above an object and another below it, without a continuous path between them while all other arms remain grasped. Such components are termed essentially mutually disconnected (EMD) when they are disconnected or cannot practically be connected within a reasonable computation time [1612.01650].

A conventional closed-chain trajectory remains within one connected component. The IK-switch provides a transition between components by temporarily releasing one arm, moving it as an open-chain manipulator to another inverse-kinematic solution for the same grasp pose, and regrasping. Environmental support, such as a floor or table, stabilizes the object during the release. The transition preserves the object-relative grasp transform while changing the internal arm configuration. Its applicability requires a collision-free open-chain transfer, a feasible support pose, static equilibrium, grasp-force feasibility, and a useful post-switch component [1612.01650].

The stability problem can be expressed through a gravito-inertial wrench and contact-force feasibility. If $\mathbf f_{\mathrm{all}}$ contains grasp and environmental contact forces, static equilibrium requires

$$
\mathbf G\mathbf f_{\mathrm{all}}=\mathbf w_{\mathrm{GI}},
$$

together with friction-cone and gripper-force inequalities. This formulation permits placements that are not passively stable under the center-of-mass support-polygon criterion when the remaining arms contribute forces and moments [1612.01650]. The associated chair-frame experiment used two six-degree-of-freedom Denso VS060 manipulators and demonstrated a hardware-executed flip whose start and goal configurations were practically EMD. The planner was run 50 times, with an average total planning time of $15.81$ s and successful compliant execution on the real platform [1612.01650].

## 2. Planning representations and discrete manipulation modes

Multi-arm planning commonly combines continuous configuration spaces with discrete manipulation modes. A mode may represent a grasp, stable object placement, handover, release, or support condition. The resulting task-and-motion problem must determine both a sequence of manipulation states and continuous collision-free motions between them.

“Preparatory Manipulation Planning using Automatically Determined Single and Dual Arms” [1812.03274] represents these alternatives with a super regrasp graph. Single-arm regrasp graphs contain transfer edges within an object-pose modality and transit edges between modalities. A handover graph connects left- and right-arm grasp configurations that realize the same object pose while maintaining gripper noncollision. The super graph contains single-arm, dual-arm, and mixed alternatives, allowing graph search to determine whether a task should use one arm, two arms, or repeated transitions between them.

The planner generates grasp configurations using force-closure considerations, combines them with inverse-kinematic solutions, and samples stable tabletop object poses. A graph node includes a gripper jaw width, arm configuration, object pose, collision-free state, and IK feasibility. A* search primarily minimizes the number of transfer and transit changes. Detailed continuous motions are then generated using Dynamic Domain RRT; if a selected graph edge is blocked, it is removed and the graph is searched again. This separation makes graph connectivity a high-level feasibility test rather than a guarantee that continuous motion exists [1812.03274].

A different representation integrates manipulation modes directly with continuous multi-arm roadmaps. “Anytime Multi-arm Task and Motion Planning for Pick-and-Place of Individual Objects via Handoffs” [1905.03179] uses a mode graph for picks, handoffs, placements, and terminal states, together with an implicit tensor-product roadmap

$$
G_1\times\cdots\times G_n.
$$

The product is explored without explicitly materializing its potentially enormous number of composite vertices. Its multi-modal extension of dRRT* jointly searches arm configurations, object transfers, and mode transitions. Unlike sequential task-and-motion planning, it does not commit early to one configuration for an underspecified arm; an unconstrained arm may move into a configuration that prepares a later handoff. The method produces an initial solution quickly and continues improving it, with asymptotic-optimality arguments for a fixed transition set and a path toward broader optimality by adding pick and handoff samples over time.

For high-dimensional manipulation with frequent inter-arm conflicts, “Accelerating Search-Based Planning for Multi-Robot Manipulation by Leveraging Online-Generated Experiences” [2404.00143] adapts Conflict-Based Search (CBS) and ECBS. Each arm has an implicit configuration-space graph, while a high-level constraint tree resolves vertex and edge conflicts between arms. The experience-accelerated variants xCBS and xECBS reuse paths and valid transitions from earlier low-level searches. A previously computed path is replayed until a new temporal constraint invalidates it; valid static collision checks are cached, while time-dependent inter-arm conflicts remain subject to verification.

This reuse preserves completeness and the stated bounded-suboptimality guarantees. In four-arm bin picking, xECBS achieved 96% success with a mean planning time of $4.4\pm3.4$ s, compared with 82% success and $18.4\pm16.9$ s for ECBS. In eight-arm shelf rearrangement, xECBS achieved 84% success and was approximately twice as fast as ECBS on successful runs. Experiments considered up to ten seven-degree-of-freedom Franka Panda arms [2404.00143].

## 3. Coordinated control, contact, and physical interaction

Planning alone does not ensure executable multi-arm behavior. Small calibration, geometric, or model errors can create incompatible end-effector poses and large internal forces when multiple arms rigidly grasp the same object. “Autonomous Dual-Arm Manipulation of Familiar Objects” [1811.08716] therefore adds an explicit kinematic-chain closure cost to synchronized trajectory optimization.

For two seven-degree-of-freedom arms, the relative end-effector transform is

$$
{}^{\mathrm{eef}_1}T_{\mathrm{eef}_2}
=
\left({}^{B}T_{\mathrm{eef}_1}\right)^{-1}
{}^{B}T_{\mathrm{eef}_2}.
$$

The optimizer penalizes deviations in relative translation and orientation from their initial values. Its STOMP-New transition cost includes collision, joint-limit, orientation, torque, duration, and closure terms. Closure increased trajectory-optimization runtime from $0.34\pm0.01$ s to $4.31\pm2.42$ s and reduced success from 100% to 83% in the reported 14-DOF lifting experiment. The complete perception-to-execution pipeline averaged approximately six seconds on the demonstrated system [1811.08716].

Closed-chain execution uses leader-follower compliant control. The leader tracks the planned joint trajectory by position control. The follower adds a force-dependent task-space correction based on wrist force/torque measurements and maps this correction through the follower Jacobian. This reduces internal stress caused by model mismatch while preserving the planned object motion [1612.01650].

Other systems relax or redistribute the coordination burden. “Manipulability optimization for multi-arm teleoperation” [2102.05414] lets an operator control a shared payload rather than individual arms. Each arm solves IK for its grasp pose, while one rotational degree of freedom around a handle axis is released. A local update evaluates the Yoshikawa manipulability index,

$$
m(\mathbf q)=\sqrt{\det(\mathbf J\mathbf J^T)},
$$

and selects a nearby configuration with improved manipulability. The system was evaluated with three six-degree-of-freedom UR5 arms and two seven-degree-of-freedom ABB YuMi arms. The method adds approximately $0.58\pm0.23$ ms over a baseline frame time of approximately 3 ms, but does not provide complete collision avoidance, regrasping, force control, or global optimization [2102.05414].

A shared-autonomy framework extends this idea to reconfigurable groups. “A Shared Autonomy Reconfigurable Control Framework for Telemanipulation of Multi-arm Systems” [2207.09813] supports Independent Control, Coordinated Control, and Freeze State. Independent Control maps a human hand to one or more robot end-effectors. Coordinated Control maps the hand to a virtual frame positioned among selected end-effectors and preserves their relative arrangement. Freeze State holds a robot or group stationary while the operator controls another group. Four seven-degree-of-freedom Franka Panda robots were used in a pallet-transport and assembly study. The reported mean normalized SUS score was $81.25\pm10.01$, and the second trial was completed 1.548 minutes faster on average, a 20.83% reduction [2207.09813].

Contact-rich manipulation motivates mode switching between geometric motion planning and force control. “An Optimization-Augmented Control Framework for Single and Coordinated Multi-Arm Robotic Manipulation” [2506.16555] assigns each manually specified subtask to pure optimization, pure force control, or hybrid control. KOMO generates long-horizon trajectories, while force PI control regulates contact forces through Jacobian-transpose torque commands. Hybrid control combines optimization, contact, and Cartesian impedance torques. In a simulated bimanual transport task, left- and right-arm force RMSEs were 2.80 N and 2.82 N, respectively; end-effector trajectory RMSEs were 1.7 cm and 1.6 cm, and box trajectory RMSE was 2.1 cm. The framework also demonstrated simulated four-arm reorientation, although it did not provide formal multi-arm stability or force-feasibility guarantees [2506.16555].

## 4. Perception, grasp generation, and articulated objects

Multi-arm manipulation depends on representations that coordinate contact choices with object geometry and the consequences of earlier grasps. “V-MAO: Generative Modeling for Multi-Arm Manipulation of Articulated Objects” [2111.03987] models contact selection as a conditional distribution over object points rather than predicting one deterministic grasp. A conditional variational autoencoder produces contact-probability maps from RGB-geometric point clouds. Different latent samples generate different plausible contact configurations, which are filtered by inverse kinematics, RRT-style planning, and collision checking.

The learned contact distribution is conditioned on the current scene after previous grasps. Its factorization is

$$
P(g_1,\ldots,g_N\mid\mathbf p^{(0)})
=
P_1(g_1\mid\mathbf p^{(0)})
P_2(g_2\mid\mathbf p^{(1)})
\cdots
P_N(g_N\mid\mathbf p^{(N-1)}).
$$

This permits later arms to account for changed object poses, newly created collision constraints, and altered reachability. Successful contact labels are generated through simulation interaction: neighboring contact tuples are sampled, executed using inverse kinematics, collision-aware planning, gripper control, operational-space control, and object-centric articulated control, and retained when the manipulation succeeds.

Object-centric control propagates the desired motion of one articulated part through an acyclic joint-dependency graph. For a revolute joint, the dependent part and its gripper are transformed using the relative rotation induced by the change in joint angle. Sequential, parallel, and sequential-parallel grasp strategies are supported. On six PartNet articulated objects and two robot platforms, V-MAO achieved more than 80% success with Sawyer and more than 70% with Panda in the reported conditions. Its main deterministic baseline, Top-1 Point, sometimes performed better on stereotyped laptop grasps, while generative sampling was more useful for variable plier contacts [2111.03987].

Perception can also be learned from human demonstrations at the level of interactions rather than trajectories. “Graph-Fused Vision-Language-Action for Policy Reasoning in Multi-Arm Robotic Manipulation” [2509.07957] extracts hand-object and object-object relations using entropy and mutual information, constructs temporal scene graphs, and converts them into language-conditioned behavior trees and Cartesian motion primitives. The graph distinguishes Coupled-Motion, Docked, Efficient Object-Object, and Transitory Object-Object relations. A cross-arm selector assigns a grasping role using hand-to-source and hand-to-target distances, a contralateral prior, demonstration preferences, and role persistence during stable placement.

In four dual-arm block-assembly benchmarks, the reported overall grasp success was 94%, placement success 89%, and task success approximately 90%. Graph representation accuracy exceeded 95%, while subtask segmentation was approximately 93%. The method does not provide formal collision avoidance, synchronization guarantees, force modeling, or comprehensive component-wise ablations [2509.07957].

A complementary research direction treats the arms themselves as active sensors. “MAAP: Multi-Agent Active Perception for Collaborative Manipulation” [2609.21929] gives every arm a wrist RGB camera and permits arms to take either manipulation or perception roles. RAIL predicts each arm’s current role and conditions action-chunk generation on that role. Across four simulated tasks, average success increased from 56.5% with a fixed global view to 62.5% with one active wrist view, 70.0% with all wrist views, and 79.2% with all wrist views plus RAIL. In the three-arm Microwave task, role-aware control increased success from 47% to 82%. On a dual-arm real platform, MAAP+RAIL succeeded in 14 of 20 placement trials compared with 0 of 20 for fixed-view ACT [2609.21929].

## 5. Learning, teleoperation, and shared autonomy

Demonstration collection is itself a multi-arm coordination problem. “Learning Multi-Arm Manipulation Through Collaborative Teleoperation” [2012.06738] introduces Multi-Arm RoboTurk (MART), which assigns one arm to each remote operator. Smartphone motion controllers and web-based video streams allow geographically separated users to operate two- and three-arm systems concurrently. The simulator waits for a command from every phone before advancing, addressing asynchronous network latency.

The benchmark includes nearly independent two-arm cube lifting, strongly coordinated tray lifting, mixed-coordination table assembly, a handover task, and three-arm lifting and wiping. The reported results show that centralized and decentralized policies have complementary strengths. Fully decentralized HBC achieved 84.5% on multi-cube lifting but 50.0% on three-arm lifting and wiping, while centralized HBC achieved 38.5% and 83.7%, respectively. The base-residual architecture combines these biases: r-HBC uses a decentralized base and centralized residual, while rd-HBC uses a centralized base and decentralized residual. r-HBC achieved 94.0% on lifting and wiping; rd-HBC achieved 89.3% on multi-cube lifting and 86.7% on tray lifting. Table assembly and pick-place handover remained difficult, with best reported success of 10.0% and 31.3%, respectively [2012.06738].

Hybrid autonomy is also used when object geometry or deformation is difficult to model. “Dual-arm Coordinated Manipulation for Object Twisting with Human Intelligence” [2108.12007] plans the motion of one arm and teleoperates the other. The planned arm grasps a known endpoint and moves to a preparation location; the human-controlled arm locates and aligns the displaced endpoint before the opposing twist. The simulation used an ABB YuMi, an alignment tolerance of approximately $5^\circ$, and a relative twist of $90^\circ$. The work reports five simulation tests but no real-robot experiments, quantitative success rates, tracking-error comparisons, force measurements, or computation-time benchmarks [2108.12007].

More recent VLA approaches make arm assignment explicit. “MA-VLA: Multi-Arm Vision-Language-Action Model for Collaboration and Compositional Generalization” [2608.25864] decomposes a global instruction into temporally ordered atomic prompts assigned to individual arms. A unified executor jointly predicts all arm actions. Arm Shuffle permutes each arm’s state, observation, prompt, action, and identity during training, discouraging fixed role-index associations. In the reported out-of-domain benchmark, baseline methods obtained zero average success while MA-VLA obtained 13%; in-domain RoboFactory performance averaged 83.5% for two-arm tasks and 83.3% for three- and four-arm tasks. The gains from Arm Shuffle were larger for compositional generalization than for ordinary in-domain execution [2608.25864].

TAO-DA takes a different route by sharing VLM-level reasoning while separating arm-specific action generation. “TAO-DA: Towards Autonomous Operation—A Dual-Arm Vision-Language-Action Model for Coordinated Manipulation” [2609.33197] uses Dual-Arm Intent Routing to classify left-only, right-only, or dual-arm execution and Dual-Tower experts to reduce cross-arm interference. Under persistent inactive-arm disturbances, Dual-Tower achieved 100% left-arm and 95% right-arm success, compared with 25% and 20% for Single-Tower. Intent-routing accuracy was 85.37% for left-only, 83.31% for right-only, and 96.13% for dual-arm trajectories. The comparison is not parameter matched because Dual-Tower approximately doubles arm-level parameters [2609.33197].

HATS extends human-agent collaboration to four-arm demonstration collection. “HATS: A Human-Agent Teleoperation System for Multi-Arm Data Collection” [2606.16491] assigns two primary arms to a human and two assistive arms to a training-free MLLM-based agent. Voice supervision can pause, retry, stop, or return an assistive arm home. Across seven real-world tabletop tasks, HATS achieved 86.4% average success compared with 90.0% for dual-human operation, 50.7% for one-human mode switching, and 35.0% for a fully autonomous agent. HATS generated downstream demonstrations whose centralized diffusion policies achieved performance comparable to policies trained from dual-human demonstrations, while hybrid policies combining learned primary behavior with deterministic assistive primitives achieved 86.9% task-completion rate for DP and 86.8% for DP3 [2606.16491].

## 6. Task allocation, active coordination, and scalable system design

Task allocation may be discrete, reactive, learned, or integrated into a larger control architecture. “Competitive Task Allocation with Temporal Proposal Windows and Improved Artificial Potential Fields for Dual-Arm Industrial Manipulation” [2610.15763] uses YOLOv8 oriented bounding boxes, a five-frame temporal proposal window, distance-based competitive allocation, and an improved artificial potential field. Two six-degree-of-freedom Kinova Gen3 Lite arms independently evaluate candidate objects, but a 0.01 m spatial-conflict threshold prevents both from selecting the same target. A 0.005 m proposal-improvement margin determines whether a newly observed candidate replaces a stored candidate. The five-frame window achieved a mean completion time of $130.51\pm1.54$ s and average velocity of $0.119\pm0.0012$ m/s in the reported ablation, outperforming the tested one-, three-, seven-, and ten-frame alternatives on the combined practical criterion [2610.15763].

Space applications expose the need to co-design arms, contacts, locomotion, and manipulation. “Design and Validation of a Multi-Arm Relocatable Manipulator for Space Applications” [2301.09863] presents MARM, a three-limb, 18-actuated-degree-of-freedom robot whose limbs can manipulate payloads, anchor to station interfaces, and support locomotion. The computational pipeline combines sampling-based planning on contact manifolds, nonlinear trajectory optimization, and inverse-dynamics quadratic programming with contact-wrench and actuator constraints. The final prototype weighs 94.4 kg. Its SIROM interface is rated approximately for 5 kN compression and radial loads, 420 Nm axial torque, and 150 Nm bending moments.

MARM’s planning problem is not ordinary free-space motion planning. Fixed stance contacts induce a nonlinear manifold, and candidate configurations are sampled in tangent spaces and projected back onto that manifold. The inverse-dynamics QP accounts for floating-base dynamics, contact constraints, actuator torques, and contact wrenches. In the reported full-gravity optimized single-stance case, hip and knee torques remained below approximately 200 Nm and ankle torques below approximately 100 Nm. The hardware experiments were preliminary and did not demonstrate complete multi-arm tile installation or locomotion [2301.09863].

A geometric abstraction can reduce the dimensionality of cooperative control. “Cooperative Task Spaces for Multi-Arm Manipulation Control based on Similarity Transformations” [2510.26362] develops a conformal-geometric-algebra representation in which a cooperative geometric primitive is described by translation, rotation, and uniform scale. These components form a seven-dimensional similarity task related to $SIM(3)=SE(3)\rtimes\mathbb R^+$. Analytic Jacobians map the aggregate joint motion to translation, rotation, and dilation coordinates. The representation is intended for bimanual manipulators, humanoids, and multi-fingered hands, where the resulting nullspace can support secondary objectives such as posture, collision avoidance, and manipulability. The supplied material establishes the mathematical machinery but does not provide the main cooperative-task definitions, complete control laws, or empirical results [2510.26362].

## 7. Limitations, evaluation issues, and research directions

The central technical limitation of multi-arm manipulation is combinatorial coupling. Adding arms increases the number of joint variables, collision pairs, grasp assignments, handover possibilities, contact modes, and synchronization relationships. Even methods that avoid explicit composite-space construction retain exponential growth in the number of possible arm and mode combinations. Tensor-product roadmaps, implicit search, experience reuse, graph abstractions, and role-conditioned policies mitigate this growth but do not remove it.

A second limitation is the gap between kinematic feasibility and physical feasibility. A path may satisfy inverse kinematics and collision constraints while violating friction, grasp-force, torque, wrench, internal-force, or object-stability requirements. Conversely, a force controller may adapt locally but fail to maintain global object orientation or long-horizon synchronization. Current systems address this through compliant control, hybrid optimization-force control, static-equilibrium checks, inverse-dynamics QPs, or learned contact distributions, but no single approach in the cited work provides a general solution for arbitrary objects, contacts, numbers of arms, and dynamic environments.

A third limitation concerns perception and representation. Category-level shape models, temporal scene graphs, wrist-camera aggregation, role labels, atomic prompts, and active-view policies improve generalization, but they depend on calibrated sensing, task-specific predicates, demonstration coverage, or manually defined role and affordance vocabularies. Learned systems often lack formal collision avoidance, force feasibility, deadlock handling, and failure-recovery guarantees. Systems with explicit geometry often lack robust perception, semantic task understanding, or scalable grasp generation.

Evaluation is also heterogeneous. Reported metrics include success rate, planning time, path cost, joint motion, force RMSE, pose error, manipulability, graph accuracy, segmentation accuracy, task-completion rate, usability, and downstream policy performance. These quantities are not interchangeable. Several studies are simulation-only, some use small structured benchmarks, and some compare systems with unequal parameter counts, different sensory inputs, or different levels of human involvement. Results therefore support bounded conclusions about particular architectures and task regimes rather than a universal ranking of multi-arm methods.

The principal research directions represented across the literature are:

1. **Integrated task, motion, and contact planning**: combining discrete mode reasoning with continuous trajectories, grasp generation, force feasibility, and execution uncertainty.

2. **Scalable coordination architectures**: sharing high-level task semantics while using arm-specific experts, role-conditioned policies, implicit tensor roadmaps, or decentralized residual controllers.

3. **Explicit synchronization and conflict resolution**: formal temporal coordination, deadlock handling, collision-aware allocation, progress-based scheduling, and communication-aware execution.

4. **Active perception through manipulation**: treating wrist cameras and mobile arms as task-driven sensing agents and optimizing viewpoints jointly with physical actions.

5. **Multimodal and generative grasping**: representing distributions of feasible contacts, incorporating articulated-object structure, and coupling sampling with motion and force validation.

6. **Hybrid autonomy and demonstration collection**: assigning semantic, contact-rich, or uncertain subtasks to humans while delegating structured or repetitive behavior to agents and deterministic primitives.

7. **Whole-body and multi-contact design**: co-designing limb morphology, actuator sizing, contact interfaces, locomotion, manipulation, and inverse dynamics for systems operating under severe workspace or support constraints.

8. **Formal guarantees under realistic execution**: extending completeness, bounded suboptimality, and stability results to continuous-time collision checking, uncertain perception, compliant contacts, dynamic obstacles, and changing numbers of manipulators.

Multi-arm manipulation is consequently best understood as a family of coupled problems rather than a single algorithmic task. Its mature formulations combine object-level representations, configuration-space reasoning, contact mechanics, role allocation, synchronized control, perception, and human or learned task-level intelligence. The most general systems are moving toward a layered design in which shared semantic reasoning determines what the team should accomplish, explicit or learned role mechanisms determine which arm should do what, geometric and dynamic modules ensure feasibility, and compliant execution adapts the plan to physical interaction.

Source: https://www.emergentmind.com/topics/multi-arm-manipulation