Grasp Planning for Robotic Manipulation

1. Introduction

Robots increasingly operate in unstructured human environments where objects rarely present themselves in configurations amenable to direct grasping. A mug wedged against a wall, a coin flat on a table, a long elastic cable that must be coiled into a narrow box, a heavy workpiece that no single arm can lift alone. Each of these scenes demands a preparatory action, a push, a slide, a tilt, a regrasp, a handover, or an explicit assignment of one arm to hold while the other acts. These auxiliary actions are not the task itself. They are the enabling moves that make the task achievable. After a decade of rapid progress on single-shot grasp synthesis for isolated objects [72, 71], the computational orchestration of such pre-manipulation actions has emerged as the dominant bottleneck for general-purpose manipulation in realistic settings.

The urgency of this problem has grown along two axes. Hardware has expanded from single-arm industrial cells to bimanual humanoids, dexterous multi-fingered hands, quadrupeds that repurpose a leg for manipulation [23], and multi-robot teams. Each additional end-effector multiplies the opportunities for enabling actions (more intermediate states, richer role assignment) while also multiplying the combinatorial difficulty of planning them. At the same time, foundation models have begun to serve as front-ends for robot planners [8, 29, 30, 31], promising zero-shot generalization to novel enabling-action scenarios. This promise sits uneasily alongside the formal and physical guarantees that classical planning has historically provided [27, 47], and the resulting tension between open-world flexibility and physical reliability now shapes much of the field's research agenda.

This survey addresses one question. What computational methods have been proposed for planning and executing auxiliary pre-manipulation actions (pushing, sliding, re-orienting, regrasping, handing off, and arm-role assignment) that enable a primary grasp when the target object is not directly graspable because of kinematic, contact, or environmental constraints, particularly in bimanual and multi-agent robotic settings. The scope is a scoping review of work from 2016 to 2026, drawing on robotic manipulation planning, bimanual and dual-arm robotics, multi-agent manipulation, task and motion planning (TAMP), and grasp planning under constraints. Foundational works from the preceding decade are included where they anchor the computational ideas that the recent literature extends [55, 56, 57, 58, 59, 69].

The field is unusually heterogeneous. It sits at the intersection of contact mechanics, combinatorial planning, reinforcement learning, imitation learning, and foundation-model reasoning. To impose structure, this survey organizes work by the principal type of enabling action addressed, mirroring the thematic decomposition adopted by contemporary reviews of nonprehensile manipulation [70] and neighboring surveys on robotic tool use. Section 3 covers pre-grasp non-prehensile manipulation (pushing, sliding, toppling to repose objects before grasping). Section 4 covers regrasping, handover, and in-hand re-orientation. Section 5 covers bimanual role assignment and coordination. Section 6 covers integrated task and motion planning frameworks that reason jointly over discrete action sequences and continuous trajectories. Section 7 covers learning-based sequential manipulation under kinematic, contact, and environmental constraints. Sections 8 through 10 examine cross-cutting themes, open problems, and the central takeaway.

2. Background and Definitions

The literature on enabling actions draws vocabulary from several communities with partially overlapping usage. This section establishes working definitions used throughout the survey and demarcates the scope.

Enabling (auxiliary pre-manipulation) actions are deliberate robot actions whose primary purpose is to transform the state of the object, the robot, or the environment so that a subsequent primary manipulation action (typically a grasp, a tool application, or a transport) becomes kinematically, dynamically, or geometrically feasible. Their value is entirely instrumental. Examples include pushing an object to a graspable pose [1, 6], regrasping to change grasp configuration [9, 10], handing an object between arms [15, 16, 17], and assigning one arm to stabilize while the other manipulates [20, 22]. The term is not universally adopted. Different communities use "pre-grasp manipulation" [63, 64, 66], "in-hand manipulation" [10], "extrinsic dexterity" [11, 41], "shared grasping" [48], or "bridge actions" to refer to overlapping phenomena. Treating these as variants of a single computational category is one of the framing choices of this survey.

Non-prehensile manipulation encompasses actions that alter an object's state without forming a stable grasp closure. These include pushing, sliding, toppling, tilting, pivoting, and dynamic transport [69, 70]. Non-prehensile actions are a principal class of enabling action because they can repose objects that are initially ungraspable, such as thin plates flat on a surface [43, 45] or large flat items in dense clutter [38, 40]. The boundary between non-prehensile manipulation as a standalone capability and as a pre-grasp enabling action is fluid. This survey attends to work that treats non-prehensile actions as means toward a downstream grasp, while the neighboring survey on robotic tool use and non-prehensile manipulation covers the non-prehensile literature whose end goal is not a grasp.

Regrasping denotes a planned change of grasp configuration on an object, potentially involving placing the object on a support surface [12, 13], transferring it between hands [9], pressing it against an environmental contact [11], or executing finger-gaiting motions [10]. In-hand re-orientation is the special case in which the object's pose relative to the hand is changed without releasing it [10, 61]. Both serve as enabling actions when the current grasp is incompatible with a required downstream manipulation. The bottleneck problem for multi-step regrasping is predicting diverse geometrically valid stable placements [12, 13].

Handover is the transfer of an object between two agents. Robot-to-robot, robot-to-human, and human-to-robot handovers have been studied, with reactive partner-adaptive policies displacing earlier static handoffs [15, 16, 17]. A handover is an enabling action when a single agent lacks the kinematic reach, dexterity, or workspace access to complete the task alone.

Bimanual role assignment is the allocation of asymmetric functional roles (holder versus actor, leader versus follower, stabilizer versus manipulator) to the two arms of a dual-arm system [18, 19, 20, 62]. Role assignment is an enabling action at the planning level. It determines which arm performs auxiliary stabilization and which executes the primary task, and empirical and computational work converges on the view that role assignment cannot be resolved independently per arm [21, 22, 23].

Task and motion planning (TAMP) is a family of frameworks that jointly reason over discrete action sequences (task planning) and continuous motion trajectories (motion planning) subject to kinematic, geometric, and increasingly dynamic constraints [24, 27, 49, 50, 56, 57]. TAMP provides a natural substrate for enabling-action planning because the need for such actions often emerges from the infeasibility of direct motion plans. Classical TAMP assumes fixed object geometry. Extending TAMP to forceful [24, 65] and deformable [25] manipulation is an active thrust.

Multi-modal motion planning treats manipulation as a search through configuration spaces that consist of intersecting submanifolds of varying dimensionality, corresponding to different contact modes [58, 59]. This formalism underlies modern treatments of grasp transitions, regrasps, and non-prehensile actions, because each contact mode change is a transition between manifolds.

Scope boundaries. This survey covers computational methods (algorithms, representations, architectures) rather than hardware design, sensing modalities per se, or human factors outside handover and shared autonomy. Manipulation of fully unconstrained objects where direct grasping suffices is excluded. Locomotion planning is excluded except where limbs are explicitly repurposed for manipulation [23]. Grasp quality metrics and force-closure analysis appear only as constraints or objectives within enabling-action planners. Deformable object manipulation is covered where enabling actions are involved [25], but not comprehensively.

3. Pre-Grasp Non-Prehensile Manipulation

When an object cannot be directly grasped because of its geometry, its pose on a support surface, or the presence of obstacles, the robot must first modify the scene. The methods developed for this purpose have evolved from hand-tuned pushing controllers in the foundational era [55, 69] to generative and learning-based policies [1, 2, 3, 39, 41] and most recently to vision-language models that select non-prehensile primitives from natural language instructions [8, 31]. The structure of the literature reflects a consistent division between the low-level question of how to execute a non-prehensile motion reliably and the high-level question of when to invoke one.

3.1 From Contact Mechanics to Learned Repositioning

The pre-deep-learning literature treated pushing and sliding as contact-mechanics problems, where friction cones, limit surfaces, and quasi-static contact assumptions yielded analytic controllers whose reliability depended on accurate parameter estimates [69, 55]. Dogar and Srinivasa [55] demonstrated that deliberate push-grasps, where the hand pushes the object rather than avoids it, extend a manipulator's reachable grasp set substantially in cluttered tabletop settings, and their framework remains one of the clearest statements of the enabling-action view of pushing. Templates for pre-grasp sliding interactions [63] and representations of pre-grasp manipulation strategies [64] established the terminology and showed that a small library of primitives (rotate-in-place, slide-to-edge, tilt-against-wall) covers a substantial fraction of household scenes. Planning pre-grasp manipulation for transport tasks [66] tied these primitives to downstream goals by demonstrating that human-inspired pre-grasp rotations and slides improve load-supporting postures and task success. This thread also produced the extrinsic-dexterity principle, where environmental contacts are exploited deliberately. Chavan-Dafle and Rodriguez [11] formalized fixtureless fixturing for regrasping, and later work [41] extended the principle to grasping otherwise ungraspable objects through emergent exploitation of walls and edges in reinforcement learning.

Beginning around 2019, end-to-end learning replaced analytic contact controllers as the default approach for non-prehensile repositioning. Haustein et al. [1] proposed a hybrid architecture that trained a GAN-based generative model to propose manipulation states, configurations from which non-prehensile actions can transport an object toward a target, coupled with a reinforcement-learning action policy for planar pushing and sliding. By treating the graspable pose as the target, this framework implicitly defined pre-grasp manipulation as goal-conditioned non-prehensile rearrangement. Cho et al. [2] extended this class of approaches to general household environments through a hierarchical and modular network that generalizes across geometries, directly addressing the narrow object-category and planar-surface restrictions of earlier systems. Del Aguila Ferrandis et al. [3] attacked a complementary weakness, the perceptual failure of vision-only pushing when the object is occluded by the gripper, by proposing a Bayesian deep-learning framework for visuotactile state estimation that supports uncertainty-aware control under occlusions after sim-to-real transfer with a simple onboard camera. Taken together, these works establish that learned non-prehensile policies can handle geometric diversity, perceptual degradation, and contact-rich dynamics. No single system yet combines all three simultaneously.

3.2 Pre-Grasp Manipulation in Clutter and Dense Scenes

Pre-grasp manipulation is particularly hard in dense clutter, where the target object is surrounded by immovable and movable obstacles and direct grasps are blocked. Learning-based retrieval policies have become the dominant response. Sun and colleagues [38] learned a pre-grasp manipulation policy that strategically pushes surrounding objects to open collision-free grasp affordances for a target, treating clutter rearrangement and grasp selection as an integrated pipeline. Liu et al. [40] specialized this idea for flat objects in cluttered scenes by learning sliding primitives that rearrange neighboring items while keeping the target pushable toward a graspable pose. Liu et al. [43] introduced a binary-mask representation of wide and flat objects for pre-grasp pushing, showing that compact shape descriptors suffice for learned pushing policies that generalize across aspect ratios. Sun et al. [44] addressed objects in ungraspable poses (large flat boxes lying on a table) by framing pre-grasp manipulation as a lift-and-secure problem solved jointly by two arms, an early demonstration that bimanual coordination is central to pre-grasp manipulation in realistic settings. Earlier work by Kim and Park [42] showed that pre-grasp manipulation can be necessary even for moderately cluttered single-object scenes when power grasps are required, because the finger's wrap-around motion needs a pocket of free space that must first be generated. The progression across these works is instructive. Early pre-grasp systems assumed that the target's surround was fixed and only the target was manipulated. The current generation treats the entire scene as actionable and learns where to act.

Analytic and hierarchical planning approaches continue to mature alongside learning. He et al. [4] introduced LDHP, a library-driven hierarchical planning framework that decouples object-motion planning (via contact-state primitives) from grasp-sequence feasibility (via gripper-aware adjustment primitives), with executability certified by quasi-static mechanics and collision constraints. This decoupling yields a training-free pipeline that transfers across object geometries and task types, directly addressing the class of thin, flat, or otherwise ungraspable objects that motivated pre-grasp manipulation in the first place. The philosophical contrast with end-to-end learning [1, 2, 3] is clear. Hierarchical decoupling sacrifices the potential for globally optimal action sequences in exchange for modularity, interpretability, and zero-shot transfer, properties that may be more valuable in safety-critical or low-data deployment scenarios. Hou et al. [45] attacked an adjacent analytic corner of the problem by showing that soft, compliant, and underactuated hands enable pre-grasp sliding manipulation of thin objects with integrated motion-and-control plans, because passive reconfigurability of the fingers absorbs contact uncertainties that rigid fingers cannot. The combined lesson is that pre-grasp manipulation is sensitive to the gripper's mechanical properties, and planners that ignore this sensitivity produce infeasible motions on real hardware.

3.3 Human-in-the-Loop and Mixed-Initiative Planning

When the search space for non-prehensile pre-grasp plans is intractably large, as in cluttered environments with many movable obstacles, autonomous planners often fail or require prohibitive computation. Two distinct mixed-initiative strategies have emerged. Papallas and Dogar [6] demonstrated that minimal human operator input integrated into a randomized control-based planner substantially reduces planning time and improves success rates for non-prehensile manipulation in clutter, establishing a human-in-the-loop planning paradigm distinct from full autonomy and direct teleoperation. Ghalamzan Esfahani et al. [7] approached the problem from the opposite direction. In a shared-autonomy teleoperation setting, an autonomous agent guided human grasp selection via haptic force cues during the reach-to-grasp phase, coupling pre-grasp pose choice to post-grasp manipulation capability through a task-relevant velocity manipulability metric (TOV). The contrast reveals the same underlying insight from two angles. Papallas and Dogar [6] use the human to reduce search complexity. Ghalamzan Esfahani et al. [7] use autonomy to improve grasp quality. Both point to mixed-initiative architectures that allocate reasoning between human and robot according to their complementary strengths. Neither work has yet been extended to dynamic, changing scenes.

3.4 Physics-Based Randomized Planning

A parallel line of research retains planning-first structure while replacing purely kinematic constraints with physics-based feasibility checks. Randomized physics-based motion planning [73] uses a simulator to evaluate proposed motions under contact dynamics, moving obstacles, and friction, so the planner can consider actions that deliberately collide with or push movable obstacles. Kinodynamic randomized rearrangement planning [53] extends this principle with nonprehensile actions that cover a wider range of dynamic transitions between statically stable states, explicitly lifting the monotonicity assumption that had restricted earlier rearrangement planners. Rearrangement planning with object-centric and robot-centric action spaces [54] showed that mixing the two action parameterizations yields stronger performance than either alone, a lesson that foreshadowed the hybrid symbolic-continuous action spaces that TAMP now relies on [49, 50]. These physics-based planners deliver the formal guarantees that learned policies lack and the efficiency that purely symbolic planners lack, but their reliance on accurate simulation remains their main liability, because friction and multi-contact dynamics are exactly the regime where simulators are least accurate.

3.5 Non-Prehensile Transport and Tray-Based Carrying

Not all pre-grasp enabling actions involve contact with the target object's surface. When the object must be moved rapidly across a workspace before it can be grasped at a new location, tray-based non-prehensile transport provides an alternative. Chen et al. [5] developed a time-optimal trajectory planner for non-prehensile 3D object transportation that exploits tray rotation as an additional degree of freedom, achieving faster transport speeds than translation-only approaches while satisfying a derived physical object-stability model and robot motion constraints. This work extends the repertoire of pre-grasp enabling actions beyond the canonical push-and-slide and exposes a subtle point. The dominant pre-grasp literature has implicitly restricted itself to planar contact actions, leaving significant capabilities (rapid non-contact transport, controlled tilting during carry, dynamic throws and catches) underexplored.

3.6 Vision-Language Models as Pre-Grasp Planners

The most recent development in this theme is the deployment of vision-language models (VLMs) as high-level planners that decide when non-prehensile primitives are needed and which to invoke. Zhu et al. [8] proposed AdaptPNP, a unified framework that uses a VLM to interpret visual scenes and task descriptions, selects between prehensile and non-prehensile action primitives (pushing, poking, sliding), and incorporates a digital-twin intermediate layer that predicts resulting object poses for mental rehearsal prior to execution. This architecture enables online replanning when execution deviates from the plan, a capability absent from the open-loop learned policies [1, 2]. Mandikal and Grauman [39] approached a different aspect of the same problem, learning dexterous manipulation from exemplar object trajectories and pre-grasps that provide dense demonstrations of how a hand should move around and across an object before the grasp forms. The two works illustrate the complementary roles VLMs and demonstration data play. The VLM supplies semantic decisions about which primitive to use. Pre-grasp demonstrations supply the continuous motion for realizing that decision. The open issue is reliability. VLM-based action selection lacks the formal guarantees of constraint-based planners [4, 27], and the digital-twin predictions that AdaptPNP relies on are only as faithful as the underlying simulator. The tension between VLM generality and physics-based reliability remains unresolved and constitutes a central open problem for pre-grasp planning.

4. Regrasping, Handover, and In-Hand Re-Orientation

Where pre-grasp non-prehensile manipulation brings an object into a graspable configuration, regrasping, handover, and in-hand re-orientation change the relationship between the object and the grasping agents after contact is established. These actions are enabling when a current grasp is incompatible with a downstream goal, when a second grasp is geometrically reachable but the first is not, or when two agents must share the object along the way. Classical treatments of humanoid regrasp planning [52] framed the problem as a motion-planning search over discrete grasp assignments and continuous arm trajectories. The past decade has reshaped this view in two ways. Graph-based contact representations unify in-hand pushing, regrasping, and hand-to-hand transfer into a single planning problem [9]. And stable-placement prediction has emerged as the principal computational bottleneck for multi-step regrasping [12, 13].

4.1 Graph-Based Representations for Grasp Reconfiguration

A unifying representational advance in in-hand manipulation planning has been the formulation of grasp reconfiguration as graph search over contact-space trajectories. Cruciani et al. [9] introduced the Dexterous Manipulation Graph (DMG), a disconnected undirected graph representing possible finger-contact trajectories on an object's surface, which unifies in-hand pushing and regrasping into a single planning problem. Critically, in a dual-arm setting, Cruciani et al. [9] demonstrated that assigning both arms interchangeable roles (neither fixed as holder nor manipulator) expands the reachable set of grasp reconfigurations, effectively treating hand-to-hand transfer as a tool within the in-hand manipulation planner rather than as a standalone handover. Sundaralingam and Hermans [10] addressed the same geometric regrasping problem through an optimization framework that alternates between finger gaiting (discrete contact relocation) and in-grasp manipulation (continuous object motion within a fixed grasp), providing a complementary analytical perspective to the graph-search formulation. Chavan-Dafle and Rodriguez [11] contributed a distinct physical mechanism, regrasping by pushing an object against external environmental contacts ("fixtureless fixturing"), which exploits extrinsic dexterity for grasp reconfiguration without requiring a second hand or dexterous fingers, broadening the applicability of regrasping to minimally actuated grippers. Shi et al. [61] explored the dynamic counterpart, using inertial loading from controlled arm motion to achieve desired in-grasp sliding, showing that regrasping need not be quasi-static. The earlier study of prehensile pushing with alternating sticking contacts [74] established the mechanical analysis underpinning the in-hand pushing trajectories that later graph-based planners search over.

4.2 Stable Placement Prediction as the Regrasping Bottleneck

For sequential pick-and-place regrasping, where an object must be placed down and re-grasped in a new configuration to reach an otherwise unreachable goal pose, the principal computational bottleneck is predicting diverse, geometrically valid stable placements on support surfaces. Xu et al. [12] made this explicit, developing point-cloud-based neural networks with orientation generation, refinement, and discrimination stages to learn stable placement prediction, and showed that this learned component is the critical enabler for multi-step regrasping pipelines toward arbitrary goal poses. Levit et al. [13] built upon this insight by interleaving the construction of a configuration-space regrasp map, an abstract model of feasible regrasp areas and grasp sequences, with an optimization-based TAMP solver, creating a feedback loop where failed motion-level refinements update the abstract regrasp model. This adaptive interleaving directly addresses a limitation of Xu et al. [12], whose pipeline assumed that placement predictions are correct at generation time without post-hoc correction. The progression from static prediction [12] to adaptive refinement [13] mirrors a broader trend in manipulation planning toward closed-loop, failure-driven replanning.

4.3 Task-Driven Regrasp Timing

A conceptually distinct question is not how to regrasp but when. At what point in a manipulation task does the current grasp become insufficient. Patankar et al. [14] formalized this by decomposing complex manipulation tasks into sequences of constant screw motions and computing graspable regions on the object's point-cloud surface for each screw segment. A regrasp is needed wherever graspable regions of contiguous segments do not overlap, transforming regrasp planning from a heuristic decision into a path-constraint-satisfaction problem where the task's kinematic structure (not a goal grasp pose) dictates regrasp timing and frequency. This task-driven formulation contrasts with the goal-pose-driven regrasping of Xu et al. [12] and Levit et al. [13], and with the contact-trajectory-driven replanning of Cruciani et al. [9] and Sundaralingam and Hermans [10], by grounding the regrasp decision in the downstream manipulation requirements rather than the current or desired grasp configuration. The two perspectives are complementary. Patankar et al. [14] determine that a regrasp is needed at a specific task phase, while graph- and placement-based methods [9, 12, 13] determine how to execute it.

4.4 Handover Planning from Static to Reactive

Robot-to-human and human-to-robot handovers have evolved from static, pre-planned motions to reactive, partner-adaptive behaviors. Göksu et al. [15] proposed a framework using Hidden Semi-Markov Models (HSMMs) to generate kinematically constrained bimanual robot motions for human-like robot-to-human handovers, learned from human demonstration data. Kim et al. [16] extended this paradigm by making the handover dynamic, where the robot adjusts its trajectory in real time to the observed receiver's motion, and provided empirical evidence from user studies that dynamic, receiver-adaptive handover significantly reduces handover time and improves user comfort compared to static strategies. Both works share the design principle of combining probabilistic motion priors (for naturalness and timing) with explicit task-space kinematic constraints (for geometric accuracy), separating perceptual naturalness from physical correctness [15, 16]. Christen et al. [17] addressed the complementary direction, human-to-robot handover, by learning control policies from point clouds in simulation, enabling vision-based handover without relying on hand-crafted heuristics for grasp timing or approach direction. A notable gap persists. While bimanual robot-to-human handover has received attention [15], bimanual human-to-robot handover and multi-agent robot-to-robot handover remain largely unexplored, despite their relevance to collaborative assembly and logistics.

4.5 Humanoid and Multi-Arm Regrasp Planning

Humanoid regrasp planning confronts the combined difficulty of high-DOF kinematics and dual-arm coordination. Vahrenkamp et al. [52] proposed efficient solutions for motion planning of dual-arm manipulation and re-grasping tasks on humanoids, showing that decomposing the planning problem into arm-level subproblems with explicit coupling constraints yields tractable plans even at high DOF. The classical approach is effective for quasi-static regrasps but does not easily handle contact-rich or dynamic motions. More recent work has added learned models at the arm level [13, 34] while retaining the dual-arm decomposition as the organizing principle. Bimanual grasp planning in the pre-2016 period [60] established the templates and force-balance criteria that newer learning-based bimanual systems [35, 37] still use as structural priors, though current practice rarely acknowledges this heritage explicitly.

5. Bimanual Role Assignment and Coordination

Bimanual manipulation is not merely two single-arm problems solved in parallel. It is a qualitatively different planning problem because the two arms share kinematic, dynamic, and force-closure constraints, and the most effective bimanual strategies assign the arms asymmetric roles that neither could play alone [51, 62]. The past decade has seen bimanual role assignment evolve from engineered skill primitives with hand-coded role labels [18, 19] to emergent role differentiation in end-to-end learning [20, 23], with analytic planning methods in between that exploit the specific structural simplifications the bimanual constraint provides [22].

5.1 The Spectrum of Role Assignment Mechanisms

Bimanual role assignment, deciding which arm holds, stabilizes, or supports while the other acts, inserts, or manipulates, has been approached through at least three fundamentally different mechanisms, reflecting divergent assumptions about what information is available at planning time. At one end of the spectrum, skill-based role encoding embeds the holder/actor distinction within reusable manipulation skill primitives. A symbolic planner sequences pre-defined skills that already encode which arm performs which asymmetric function [18, 19, 20]. This approach is pragmatically effective for known task repertoires but is inherently limited to the role decompositions anticipated by the skill designer. At the other end, emergent role differentiation arises from end-to-end learning without explicit role assignment. Arm et al. [23] demonstrated that in a quadruped repurposing one leg for manipulation, asymmetric roles (actor limb versus collective stabilizer limbs) emerge from a unified reinforcement-learning controller that adaptively modifies locomotion gait to extend the actor limb's reachable workspace. Between these extremes, Sundaram et al. [19] proposed actuation-matrix-driven role differentiation, where an explicit matrix encoding contact-force generation differences between heterogeneous hands (fully actuated versus underactuated) determines role assignment based on physical force-feasibility and manipulability constraints, grounding role allocation in mechanics rather than task semantics or learned behavior. Guiard's classic human-bimanual framework [62] anticipated the core distinction by showing that skilled human bimanual action exhibits asymmetric division of labor where the non-dominant hand stabilizes a frame of reference while the dominant hand acts within it, a template that many of the computational frameworks [18, 20, 35] recapitulate implicitly.

5.2 Coupled Planning and Bidirectional Constraint Propagation

A recurring finding across the bimanual coordination literature is that effective role assignment cannot be resolved independently for each arm. Instead, it emerges from interactively coupled planning constraints that propagate bidirectionally between the two limbs. Huhn et al. [21] provided psychophysical evidence from human bimanual manipulation showing that planning constraints couple the two hands. The holder arm's planning demands constrain the actor arm's and vice versa, with task symmetry structure shaping this coupling. Cohn et al. [22] formalized this computational insight by developing an analytic IK-based constraint manifold reduction for rigid co-manipulation. The fixed inter-end-effector transformation is resolved by an analytic inverse kinematics solution that re-parametrizes joint configuration space into a lower-dimensional representation where feasible configurations have positive measure rather than forming a measure-zero set. This re-parametrization enables sampling-based planners, trajectory optimizers, and convex inner-approximation methods to operate on the constrained bimanual problem without modification. Arm et al. [23] achieved analogous coupling through learning rather than analysis. The RL-trained controller implicitly learns gait-adaptive coordination where the supporting limbs restructure their behavior to expand the manipulation limb's workspace, recovering bidirectional constraint propagation as an emergent property of reward optimization rather than an engineered computational structure. Earlier humanoid-manipulation planners by Vahrenkamp et al. [52] established the tractability of explicit coupling with high-DOF humanoids, but did so using hand-engineered constraints, so the theoretical progress across these works is best understood as a move from explicit to implicit encoding of the bimanual coupling while preserving its bidirectional character.

5.3 Reactive Role Alternation

Classical bimanual coordination assumes a fixed temporal ordering, where one arm stabilizes, then the other acts. Realistic manipulation tasks often require reactive alternation between roles. Grannen et al. [20] introduced a learned binary classifier that continuously monitors whether the environment has drifted from a stable configuration during task execution, triggering the holder arm to re-establish its grasp before the actor arm resumes. This enables closed-loop, bidirectional alternation between stabilization and action, with the restabilization condition itself learned from demonstrations rather than analytically specified. The shift from sequential hand-off [18] to reactive alternation [20] represents an important maturation of bimanual coordination, though it introduces new challenges. The stability classifier must generalize across object geometries and task contexts, and false positives (unnecessary restabilization) degrade task efficiency while false negatives (missed instability) risk catastrophic failure. Sensor-conditioned skill transitions using multi-modal sensing (vision, force/torque, tactile) have been proposed to gate transitions between bimanual skills [18, 20], but robust, generalizable transition monitoring remains an open problem.

5.4 Bimanual Handling of Ungraspable Objects

A distinctive class of bimanual enabling actions exists for objects that no single hand can grasp. Sun et al. [44] demonstrated that large flat boxes lying on a table, which provide no graspable edge for a single hand, can be manipulated by a pipeline where one arm lifts the box edge while the other slides underneath. The two-arm lift-and-grasp separates into distinct roles with an explicit handover of the supporting function from one arm to the other. Cai et al. [35] extended this approach to objects under changing external forces, showing that sampling stable intersections in grasp manifolds enables efficient regrasp-free transitions between uni-manual and bi-manual grasps. The grasp manifold formalism recasts the bimanual-versus-unimanual choice as a smooth trajectory through configuration space rather than a discrete decision. The classical bimanual grasp-planning literature [60] provides the geometric foundations that these learning-based systems rely on, because bimanual grasps must simultaneously achieve force closure for the object and collision-free kinematics for both arms, and the feasibility conditions have been characterized analytically for parallel-jaw and multi-finger hands. The dual-arm survey by Smith et al. [51] catalogs the coordination paradigms (position-controlled master-slave, hybrid force/position, impedance-based leader-follower) that current learning-based bimanual systems rediscover with neural networks.

6. Integrated Task and Motion Planning for Enabling Actions

Task and motion planning is the natural substrate for enabling-action reasoning because the need for such actions arises precisely when direct motion planning fails and a symbolic detour (place down, regrasp, push aside, hand over) is required. The classical TAMP foundations established in the previous decade [56, 57, 58, 59] treated the problem as search over alternating discrete and continuous decisions under kinematic and geometric constraints, and the canonical algorithmic contributions of that era (hierarchical planning in the now [56], factored task-and-motion planning [49, 50], logic-geometric programming [50]) remain the backbone of contemporary systems. The developments of the past decade extend TAMP along four axes. Force and deformation become first-class constraints [24, 25, 46, 47, 48, 65]. Enabling actions are discovered by graph repair rather than pre-specified [26]. Formal correctness certificates complement probabilistic guarantees [27, 28]. And foundation models serve as perception and task-decomposition front-ends [29, 30, 31].

6.1 Extending TAMP Beyond Kinematics

The classical TAMP formulation reasons over kinematic and geometric constraints, including collision avoidance, reachability, and stable placement. Many enabling actions involve forces (prying, twisting, pressing) or deformable objects whose geometry changes under manipulation. Holladay et al. [24] extended TAMP to forceful manipulation by augmenting the framework with wrench controllers and torque/friction limit constraints, making force feasibility a first-class constraint driving both discrete strategy selection and continuous parameter optimization. Critically, Holladay et al. [24] also proposed robustness-to-parameter-variation as a criterion for strategy selection. Strategies are preferred when their required continuous parameters are insensitive to perturbations, coupling robustness analysis to symbolic action sequencing, a departure from the purely feasibility-driven selection of earlier TAMP solvers. Toussaint et al. [47] had earlier provided the theoretical grounding by formulating force-based sequential manipulation as an optimization over physics-based feasibility constraints, demonstrating that force considerations can be embedded in logic-geometric programming rather than delegated to a post-hoc controller. Ma et al. [25] addressed a complementary extension by developing geometric-state-aware action planning for elastic objects (packing long linear elastic objects into boxes), where the discrete action sequence must account for the object's evolving shape after each manipulation step using a hybrid geometric model. Chen and Berenson [46] and Chen and Berenson [48] extended the notion of environmental contact from an obstacle to be avoided to a resource to be exploited. Manipulation planning that uses environmental contacts to keep objects stable under external forces [46] and the shared-grasping framework that trades hand contacts for environment contacts [48] expand the enabling-action repertoire by giving the planner explicit means to recruit walls, edges, and fixed surfaces as temporary stabilizers. Cheng et al. [65] brought force-and-motion constrained planning to tool use, showing that enabling-action planning for tool applications must handle continuous force application over whole trajectories, not just at contact transitions.

6.2 Enabling Action Discovery as Graph Repair

A powerful framing for enabling-action planning casts the problem as graph repair. When adjacent action primitives in a task graph have incompatible state interfaces, the planner automatically retrieves and splices in an intermediate enabling action. Takata et al. [26] instantiated this idea in a bimanual cooking domain, decomposing recipes into task graphs and inserting intermediate motions from a database when adjacent primitives require bridging. This graph-connectivity-driven approach also yields bimanual parallel task execution as a structural consequence. Simultaneous dual-arm scheduling emerges by identifying independent sub-chains in the graph and assigning each to one arm, making parallelism a property of graph topology rather than a separately solved coordination problem [26]. Garrett et al. [49] provided the more general algorithmic mechanism with PDDLStream, which combines symbolic planning with blackbox samplers so that continuous parameters (where to push, which grasp to use, which placement to select) are sampled adaptively against symbolic preconditions and effects. PDDLStream has become a workhorse for enabling-action planning because it makes continuous auxiliary actions first-class in a PDDL-style formulation without requiring bespoke integration code. Logic-geometric programming [50] remains the optimization-based counterpart for scenarios where the cost function over final geometric states is the primary objective, and the two paradigms increasingly coexist in modern systems.

6.3 Formal Guarantees in Reactive TAMP

The graph-repair and sampling-based approaches provide probabilistic or heuristic guarantees. A complementary thread in the TAMP literature develops architectures with provable correctness. Vasilopoulos et al. [27] encoded complex manipulation tasks as linear temporal logic (LTL) formulas enriched with mobile manipulation primitives, compiling them to automata that reactively ground discrete plan states to continuous controllers online. The key contribution is the provision of provable discrete completeness and continuous termination guarantees, introducing formal correctness certificates as a first-class property of TAMP, in contrast to the heuristic or probabilistic guarantees of graph-repair methods [26, 49]. Vasilopoulos et al. [28] extended this architecture with a hierarchical deliberative-reactive decomposition that combines domain-independent sampling-based deliberative planning with a global reactive planner, providing further robustness in partially known environments. The trade-off these systems make explicit is the cost of formal guarantees in modeling effort. LTL and automaton-based systems require careful specification, and their applicability to contact-rich manipulation where the specifications themselves are hard to write remains a barrier that learning-based TAMP architectures [37] sidestep at the cost of the guarantees.

6.4 Foundation Models as TAMP Front-Ends

The integration of vision-language models and large language models as front-end grounding modules for TAMP has emerged as a prominent trend since 2024, enabling zero-shot generalization to open-world environments without task-specific training or pre-specified symbolic domain models. Tang et al. [29] proposed a VLM-based scenario grounding stage for humanoid bimanual dexterous manipulation, where the VLM resolves unstructured natural-language instructions against the current visual scene before downstream planning. Schakkal et al. [30] developed a hierarchical vision-language planning framework for multi-step humanoid manipulation that uses VLMs for both scene understanding and plan generation. Lee et al. [31] investigated LLM-based planning for non-prehensile tool-object manipulation in confined environments, combining the language model's task decomposition with a manoeuvrability-driven stepping controller that uses visual-feedback tool affordance models to define geometric reachability for tool-object interactions. The convergence across these works is striking. Foundation models serve as translators between unstructured task specifications and structured TAMP inputs, rather than as end-to-end controllers. This division of labor introduces a brittle interface. The foundation model may ground tasks in physically infeasible ways that the downstream TAMP solver cannot recover from, and none of these works provide formal guarantees comparable to Vasilopoulos et al. [27]. The reliability gap between foundation-model-driven and formally grounded TAMP remains a central unresolved tension.

6.5 Rearrangement Planning as a TAMP Stress Test

Rearrangement planning, where several movable objects must be relocated so a target can be reached, is a stress test for TAMP because it combines combinatorial discrete choice (which objects to move) with continuous motion (how to move them) under physical feasibility constraints. King et al. [67] and Dogar and Srinivasa [55] established the tractability of rearrangement planning with non-monotone plans, where an object may be moved, then moved again, before the goal state is reached. Krug et al. [53] demonstrated kinodynamic randomized rearrangement with dynamic transitions between statically stable states, lifting the quasi-static assumption. Stilman et al. [68] provided probabilistic completeness results for manipulation among movable obstacles that continue to inform contemporary planners, and Hertle and Nebel [54] demonstrated that mixing object-centric and robot-centric action parameterizations yields stronger performance than either alone. The current generation of learning-augmented rearrangement planners [13, 37] inherits this discrete-continuous structure and adds learned samplers for continuous parameters and learned heuristics for discrete choice.

7. Learning-Based Sequential Manipulation Under Constraints

Learning-based methods now account for a growing share of enabling-action research, particularly where the contact dynamics, the perceptual bottlenecks, or the combinatorial scale defeat analytic planners. This section organizes the learning literature into four closely related threads. Reinforcement learning that exploits extrinsic dexterity to overcome kinematic constraints [32, 33, 41]. Multimodal policy representations for contact-rich manipulation [32, 35, 36]. Hybrid learning-planning architectures for long-horizon bimanual tasks [37]. And data-efficient demonstration pipelines that sidestep sim-to-real transfer for enabling-action skills [34].

7.1 Exploiting Extrinsic Dexterity Through Reinforcement Learning

A key insight driving recent RL-based approaches to pre-grasp manipulation is that environmental contacts (table edges, walls, other objects) can be deliberately exploited as extrinsic dexterity to overcome the kinematic limitations of the robot's own hand. Wu et al. [32] demonstrated this by learning dexterous functional pre-grasp manipulation via RL, where a dexterous hand repositions and re-orients objects by leveraging surface contacts, with a mixture-of-experts strategy followed by diffusion policy distillation to capture the multimodal action distributions arising from diverse valid solutions under contact constraints. Pavlichenko and Behnke [33] tackled a closely related problem, learning pre-grasp manipulation for human-like functional grasps, but introduced abstract, constraint-based grasp representations (specifying relational and geometric constraints rather than explicit target poses) as learning targets, enabling functional grasp achievement without full pose supervision. Zhou and Held [41] provided an early and influential instance of the extrinsic-dexterity principle by showing that a simple gripper can emergently learn to push objects against walls or edges to regrasp, without any hand-coded contact assumptions, when trained end-to-end in a simulation that includes such contacts. The shift from explicit pose targets [1] to abstract grasp constraints [33] parallels the shift from goal-pose-driven to task-driven regrasping in the regrasping literature [14], suggesting a broader methodological trend toward constraint-centric rather than configuration-centric manipulation planning.

7.2 Multimodal Policy Representations for Contact-Rich Manipulation

Contact-rich manipulation under constraints generates inherently multimodal action distributions. Multiple distinct strategies (different push directions, alternative finger placements) may be equally valid for achieving the same pre-manipulation goal. Wu et al. [32] showed that diffusion policy distillation from a mixture of experts outperforms unimodal policy representations for dexterous pre-grasp manipulation, directly addressing the mode-collapse problem that afflicts standard policy-gradient methods in contact-rich settings. Cai et al. [35] extended multimodal reasoning to the bimanual setting, demonstrating that sampling stable intersections in grasp manifolds enables efficient, regrasp-free transitions between uni-manual and bi-manual grasps under varying external forces, paired with a hierarchical imitation learning global planner and QP-based local planner that balances long-horizon path feasibility with real-time constraint satisfaction. Zhang et al. [36] addressed the coordination dimension of multimodality. Attention-based intrinsic regularization (DAIR) in bimanual RL enforces sub-task specialization between cooperating arms, preventing both domination (one arm monopolizing the task) and workspace conflict (inter-arm collision), yielding safer and more efficient cooperative policies under sparse rewards without explicit collision penalty terms. These three works collectively demonstrate that multimodal representations are not merely a technical refinement but a fundamental requirement for learning manipulation policies in constrained, multi-contact, multi-agent settings.

7.3 Bridging Learning and Planning for Long-Horizon Bimanual Tasks

Pure learning and pure planning approaches each exhibit characteristic failure modes for long-horizon bimanual manipulation. Learned policies suffer from error accumulation and poor out-of-distribution generalization. Classical planners struggle with perceptually complex sub-tasks and require extensive domain engineering. Recent work has sought hybrid architectures that allocate reasoning between learned and planned components according to their strengths. Chen et al. [37] proposed SViP, integrating visuomotor imitation learning into TAMP via a semantic scene graph monitor that detects task-phase transitions and generates parameterized scripted primitives as fallback bridges. The TAMP constraint formalism handles out-of-distribution recovery and unseen task generalization, while learned visuomotor policies handle perceptually complex in-distribution sub-tasks, achieving strong generalization from as few as 20 real-world demonstrations without object pose estimators. The complementary role played by Levit et al. [13] in the regrasp setting (using learned models to feed a constraint-based TAMP solver while allowing the planner to override failed learned predictions) illustrates the same architectural pattern from a planner-centric rather than learning-centric perspective. Taken together, these systems suggest a maturing consensus. Enabling-action planning benefits from a learned perception and control layer, a classical planning layer for sequencing and constraint satisfaction, and a monitoring layer that mediates handoffs between them.

7.4 Data-Efficient Demonstration Pipelines

A complementary bottleneck for learning-based enabling-action methods is data acquisition. Contact-rich, bimanual, long-horizon demonstrations are expensive to collect, and teleoperation is awkward for whole-body humanoid manipulation. Nai et al. [34] addressed this by proposing robot-free whole-body demonstration collection via portable wearable motion capture, paired with a hierarchical imitation learning pipeline that explicitly bridges human-robot embodiment differences, producing diverse humanoid manipulation skills that generalize to unseen environments without teleoperation hardware or complex RL reward engineering. The complementarity with Chen et al. [37] is notable. Chen et al. solves the execution architecture problem (how to combine learned and planned components). Nai et al. solves the data acquisition problem (how to collect diverse demonstrations efficiently). Neither alone is sufficient for deployable long-horizon bimanual manipulation, but their combination points toward a viable pipeline. Mandikal and Grauman [39] occupy an earlier position on the same axis, using exemplar object trajectories and pre-grasps extracted from human video as a lighter-weight source of supervision for dexterous manipulation. The central question that remains open across these works is the quality of the embodiment transfer. Human motion-capture data and exemplar human videos contain biomechanical constraints that differ from the robot's, and the hierarchical bridges these pipelines use are the current best attempt to absorb the mismatch without resorting to explicit retargeting networks.

8. Cross-Cutting Analysis

Looking across the five themes, five methodological axes recur and illuminate both the convergences and the unresolved tensions in the field.

8.1 The Learning-Planning Continuum

The most salient methodological axis across all five themes is the continuum between purely learned and purely planned approaches to enabling actions. At the planning end, constraint-based methods provide interpretability, formal guarantees, and zero-shot transfer [4, 22, 27, 28, 49, 50]. At the learning end, RL and imitation learning methods handle perceptual complexity, contact-rich dynamics, and multimodal action distributions [2, 3, 32, 36]. The most productive recent systems occupy the middle of this continuum. Learned components handle perception and low-level control, while planning components handle sequencing and constraint satisfaction [8, 13, 37]. This hybrid architecture is not simply a pragmatic compromise. It reflects a fundamental decomposition of the enabling-action problem into sub-problems with different computational structures (continuous optimization versus discrete search, data-driven versus model-based).

8.2 The Foundation Model Inflection Point

The period 2024 to 2026 has seen a rapid inflection in the use of VLMs and LLMs as planning front-ends [8, 29, 30, 31]. These models provide unprecedented natural-language grounding and zero-shot generalization, but uniformly lack the physical reasoning capabilities and formal guarantees of classical TAMP [24, 27]. The emerging design pattern (foundation model as scene interpreter and task decomposer, TAMP or learned controller as executor) recapitulates the classical AI architecture of perception-planning-execution with learned rather than engineered perception. Whether foundation models can be endowed with sufficient physical common sense to serve as reliable enabling-action planners, or whether they will remain limited to high-level task decomposition, is arguably the most consequential open question for the next decade.

8.3 Bimanual and Multi-Agent Coordination as an Amplifier

Bimanual and multi-agent settings do not merely add complexity. They qualitatively expand the space of possible enabling actions by enabling handovers [15, 16, 17], inter-arm stabilization [20], and role-based task decomposition [18, 19, 23]. The coordination overhead is substantial, but the literature consistently demonstrates that bimanual systems can accomplish tasks impossible for single arms. Manipulating ungraspable objects via dual-arm regrasping [9, 44]. Co-manipulating constrained objects via coupled planning [22, 52]. Achieving parallel task execution through graph-structural analysis [26, 29]. A notable convergence is the shift from pre-specified role assignment [18] to emergent or adaptive roles [20, 23], mirroring the broader learning-over-engineering trend. An unresolved question is whether the coordination principles discovered in bimanual settings will extend cleanly to three-arm, quadrupedal-leg, or multi-robot settings, where the combinatorics of role assignment become qualitatively harder.

8.4 Sim-to-Real Transfer and Evaluation Gaps

A methodological concern that cuts across all themes is the predominance of simulation-only evaluation. Many of the learning-based systems reviewed [2, 3, 36, 41] report results primarily in simulation, with limited or no real-robot validation. Sim-to-real transfer is particularly challenging for enabling actions because they involve contact-rich interactions where simulation fidelity is lowest (friction, deformation, multi-contact dynamics). The works that do demonstrate real-robot transfer, notably Chen et al. [37] with 20 demonstrations, Nai et al. [34] with wearable motion capture, and Kim et al. [16] with user studies, provide significantly stronger evidence, but their evaluation is typically limited to a small number of tasks and objects. Systematic benchmarking of enabling-action methods across diverse objects, constraints, and platforms remains a significant gap. The neighboring non-prehensile survey [70] and the data-driven grasp synthesis survey [72] provide template benchmarks (planar pushing, grasp quality on standard object sets) that enabling-action research has not yet adopted in standardized form.

8.5 From Configuration-Centric to Constraint-Centric Planning

A subtle but important methodological shift visible across multiple themes is the move from configuration-centric to constraint-centric representations of enabling actions. Early regrasping work specified target grasp poses as 6-DOF transforms and searched for motion plans to reach them [52, 60]. Recent regrasping work specifies relational and geometric constraints and treats the explicit pose as an inferred quantity [14, 33]. Early pre-grasp pushing specified target object poses [1, 43]. Recent pre-grasp pushing specifies graspability constraints [38, 40] or abstract relations between objects and fingers [33]. Early TAMP specified goal configurations for all movable objects [56, 57]. Contemporary TAMP specifies temporal-logic properties [27] or natural-language goals grounded by VLMs [29, 30]. The constraint-centric formulation has two advantages. It better matches the structure of human task descriptions, which are typically relational rather than metric. And it leaves more freedom for planners and policies to discover solutions, because any configuration satisfying the constraints is acceptable. The trade-off is that constraint satisfaction can be hard to verify, and a constraint that is too loose yields incoherent behavior.

8.6 Taxonomy of Enabling Actions

9. Open Problems and Future Directions

Six specific open problems emerge from the cross-cutting analysis and deserve targeted attention from the next generation of enabling-action research.

Unified enabling-action taxonomies and benchmarks. The field lacks a shared taxonomy of enabling actions and standardized benchmarks for evaluating them. Section 8.6 presents one candidate taxonomy, but no equivalent exists in the literature as a consensus structure. Future work should develop a principled taxonomy grounded in the type of constraint the enabling action resolves (kinematic, contact, environmental, force) and create benchmark suites with diverse objects, constraints, and success metrics that span the themes identified in this survey. Without such benchmarks, cross-method comparison remains unreliable. The template set by the data-driven grasp synthesis survey [72] and the nonprehensile dynamic manipulation survey [70] offers a model for the kind of shared evaluation protocol the enabling-action community could adopt.

Closing the VLM-physics gap. Foundation-model-based planners [8, 29, 31] generalize impressively but lack physical grounding. A promising direction is neurosymbolic TAMP that uses VLMs for task decomposition and scene grounding while enforcing physics-based feasibility constraints through classical mechanics models, potentially combining the formal guarantees of Vasilopoulos et al. [27] with the open-world generalization of foundation models. The digital-twin component of AdaptPNP [8] is an early step in this direction but does not provide the guarantees that LTL-based reactive TAMP offers. Formal synthesis of LLM-produced action sequences against constraint specifications is an underexplored avenue.

Multi-step enabling-action chains. Most current methods plan at most one or two enabling actions before the primary manipulation. Real-world tasks may require chains of enabling actions (push, regrasp, hand over, re-orient) whose planning requires joint reasoning across the theme boundaries identified in this survey. Levit et al. [13] and Chen et al. [37] represent initial steps. The classical multi-modal planners [58, 59] and their rearrangement successors [53, 67, 68] handle chains in principle, but their computational cost grows quickly with chain length, and no current system handles a realistic kitchen or assembly workflow end to end. Search-guidance from foundation models may be the most viable path to scale.

Contact-rich sim-to-real for bimanual systems. The sim-to-real gap for contact-rich bimanual enabling actions (dual-arm regrasping, coordinated pushing, handover under deformation) requires targeted investigation. Domain randomization and teacher-student distillation [3, 32] have shown promise for single-arm contact, but bimanual contact multiplies the fidelity requirements. Real-robot demonstration pipelines like Nai et al. [34] may offer a complementary path by sidestepping simulation entirely for data collection, and the quality of the embodiment-transfer bridge remains the main open variable there.

Formal verification of learned enabling-action policies. As learned policies replace planned enabling actions, the formal guarantees provided by systems like Vasilopoulos et al. [27] are lost. Developing methods for verifying or certifying the safety and completeness of learned enabling-action controllers, particularly in bimanual settings where inter-arm collision is a critical safety constraint [36], is essential for deployment in human-proximate environments. The existing formal hybrid-systems machinery is applicable in principle, and there is scope for extending reachability analysis to learned neural controllers constrained within a TAMP envelope.

Adaptive role assignment under uncertainty. Current bimanual role-assignment mechanisms are either pre-specified [18, 19], learned offline [23], or reactively triggered by simple classifiers [20]. Developing role-assignment policies that adapt online to uncertain object properties (mass, friction, deformability), dynamic task requirements, and partial observability would significantly expand the applicability of bimanual enabling-action systems. Extending Guiard's asymmetric-division framework [62] to computational role-assignment policies remains an open research thread, particularly for many-arm systems where the taxonomy of roles itself must be learned or inferred rather than pre-specified.

Deformable and articulated targets as first-class citizens. The current enabling-action taxonomy is implicitly rigid-body-centric. Ma et al. [25] provides one of the few treatments of deformable-object enabling actions, and Chen and Berenson [46] considers external forces on otherwise rigid objects. Articulated objects, cables, fabrics, and soft food items require new primitives whose state space is richer than pose and whose contact dynamics defy standard simulation. The neighboring non-prehensile manipulation literature [70] provides partial scaffolding, but the specific pre-grasp, regrasp, and handover challenges for deformables remain open.

10. Conclusion

Over the past decade, computational methods for planning and executing auxiliary pre-manipulation actions have evolved from narrow, single-primitive systems to increasingly integrated architectures that combine non-prehensile repositioning, regrasping, handover, role assignment, and sequential planning under kinematic, contact, and environmental constraints. The central methodological tension between learning-based approaches that handle perceptual complexity and contact dynamics, and planning-based approaches that provide guarantees and interpretability, has proven productive, with the most capable recent systems occupying a hybrid middle ground. Foundation models have opened the door to open-world generalization but have not yet earned the trust that formal or physics-based methods command.

Three observations deserve emphasis. First, the problem has not become simpler with better hardware. More arms and richer end-effectors expand the enabling-action repertoire and the combinatorial planning burden in tandem, and the bimanual and multi-agent settings remain the more faithful image of the problems future robots will face. Second, the representational ground has shifted. Configuration targets have given way to relational and constraint-based targets, and classical pose-centric planners are being complemented by semantic graph monitors, regrasp maps, and foundation-model grounders that operate on richer descriptions of the task. Third, the reliability questions that early TAMP systems answered with formal guarantees are being reopened by learning-based and foundation-model-based systems that cannot offer the same assurances, and closing this gap is the single most consequential research problem for the coming years.

The single most important takeaway from this survey is that enabling actions are not peripheral to manipulation but central. The ability to plan and execute the preparatory actions that make grasping possible is, in many real-world settings, the harder and more consequential computational problem than the grasp itself. Grasp planning, in its mature form, is a sequential decision problem about the pushes, slides, regrasps, handovers, and role assignments that must occur before a primary grasp becomes feasible.