Robotic Tool Use and Non-Prehensile Manipulation

1. Introduction

The ability to alter an object's state without forming a stable grasp, pushing, sliding, pivoting, toppling, and the ability to wield tools to extend physical reach and mechanical advantage, are among the capabilities that most clearly separate versatile manipulators from pick-and-place automata. Decades of robotics research have yielded increasingly capable grasp planners and motion generators, yet the vast majority of real-world manipulation tasks demand strategies that fall outside the prehensile paradigm. A warehouse robot must slide a box too large to pinch, a kitchen assistant must flip food with a spatula rather than a finger, and a maintenance platform must press a tool against a workpiece with regulated force. These scenarios call for non-prehensile manipulation, robotic tool use, or, most demandingly, both at once.

The convergence of several technical developments makes a joint survey of these two research threads timely. Contact-implicit trajectory optimization has matured to the point where it offers a principled alternative to combinatorial contact-mode enumeration, with state-triggered and variable-smooth formulations closing much of the gap between fidelity and tractability [36, 4, 5]. Deep reinforcement learning now achieves zero-shot sim-to-real transfer for contact-rich manipulation [38, 42, 44]. Affordance learning has evolved from single-object feature classifiers into relational, multi-granularity models that capture the asymmetric structure of tool-target interactions [22, 26, 28]. And vision-language models have begun to bridge the gap between natural-language intent and low-level non-prehensile control [3, 39]. Yet despite this momentum, most existing surveys treat non-prehensile manipulation [9, 56] and tool use [57, 58] as separate problems, leaving the field without a unified map of the computational methods, representations, and learning paradigms that span both capabilities.

This review asks a single question. What are the computational methods, representations, and learning paradigms that enable robotic tool use and non-prehensile manipulation, and how do these two capabilities intersect to extend the range of tasks robots can perform? We adopt a scoping methodology covering 2014 to 2026, with foundational works from earlier years included where they establish essential theoretical grounding. The review draws on literature from robotic manipulation, planning and control, robot learning, and cognitive robotics.

The remainder of the survey is organized as follows. Section 2 establishes definitions and scope boundaries. Sections 3 through 6 present four thematic analyses. Section 3 addresses model-based planning and control for non-prehensile manipulation. Section 4 treats tool use reasoning and affordance learning. Section 5 surveys learning-based approaches, spanning reinforcement learning, imitation learning, and sim-to-real transfer. Section 6 examines integrated systems where tools extend non-prehensile capabilities. Section 7 provides a cross-cutting analysis identifying convergent trends and methodological shifts. Section 8 maps specific open problems. Section 9 concludes with a synthesis of the state of knowledge and the single most important direction for future research.

2. Background and Definitions

Non-prehensile manipulation refers to any manipulation strategy that alters an object's state without establishing a form-closure or force-closure grasp. It encompasses pushing (translating an object through sustained frictional contact), sliding (controlled object motion across a supporting surface), pivoting (rotating an object about a contact point with the environment), toppling (inducing controlled tipping), tossing (ballistic launching), and non-prehensile transport (carrying an object on a tray or palm without enclosure). The defining characteristic is that the manipulator does not kinematically constrain all degrees of freedom of the object. Environmental contacts, gravity, and friction contribute to the object's motion and stability [10, 1]. The analytical foundations trace to Mason's seminal work on planar pushing mechanics [8], which established the quasi-static framework that remains the basis for most modern contact models, and to Lynch and Mason [47], who proved controllability of the pusher-slider system and derived stable pushing strategies that maintain predictable contact. The resulting mismatch between the robot's fully actuated kinematics and the object's partially constrained dynamics is both the source of non-prehensile manipulation's versatility and its central computational challenge.

Robotic tool use refers to the purposeful wielding of an intermediary object, the tool, to achieve a manipulation objective that would be difficult or impossible with the robot's bare end-effector. Tool use subsumes grasping the tool, reasoning about its functional properties, planning how to bring it into effective contact with a target, and executing the resulting contact-rich motion [46, 45]. Tool use is distinct from fixture-based manipulation, where the environment geometry is passively exploited, in that the tool is actively grasped and controlled by the robot. It is also distinct from end-effector design, although the boundary blurs when modular tool attachments are swapped during task execution [1].

Affordances, following Gibson's ecological psychology and its robotics formalization by Şahin et al. [46], denote the action possibilities that an environment offers an agent. In the context of this review, an affordance model maps from perceptual features of objects (geometry, material properties, spatial relations) and agent capabilities (reach, grasp repertoire, force capacity) to predicted action outcomes. Affordance models range from hand-coded functional features [49] to learned probabilistic mappings [48, 23] to structured knowledge graphs [28]. A key distinction is between single-object affordances (what can I do with this object?) and relational affordances (what can I do to object B using object A as a tool?). The latter are essential for tool use [52] and have emerged as the dominant paradigm over the review period.

Scope boundaries. This review does not cover purely prehensile manipulation (grasp planning, in-hand reorientation without environmental contacts), prosthetic or soft robotic tool use, tool manufacturing or 3D printing of tools (except where the design process is learned as a policy [17]), or cognitive-science studies of animal or human tool use except where they directly motivate computational models. We focus on computational methods (algorithms, representations, and learning frameworks) rather than mechanical design of end-effectors or tools, although we note hardware developments where they are tightly coupled with the algorithmic contribution.

The survey organizes roughly seventy papers around an expanded thematic taxonomy. A compact view of the four themes and the representative classes of methods under each appears below.

3. Planning and Control for Non-Prehensile Manipulation

3.1 The Randomized Planning Paradigm and Its Computational Limits

Control-based randomized planning served for much of the review period as the dominant paradigm for non-prehensile manipulation in clutter, yet it faces persistent computational bottlenecks that have motivated over a decade of algorithmic improvement. Approaches in this family typically interleave random configuration sampling with physics-based forward simulation, and they suffer from planning times on the order of tens of seconds to minutes and success rates that degrade sharply on difficult instances [30, 6, 19]. The core difficulty is the high dimensionality of the joint robot-object configuration space coupled with the need to reason about frictional contact modes at every candidate interaction. Muhayyuddin et al. [59] and Agboh and Dogar [60] showed that physics-based sampling under state uncertainty can be sped up by caching rollouts and re-using them across online replanning, but even these optimizations leave the underlying combinatorial structure intact. Bejjani et al. [61] proposed receding-horizon planning with a learned value function to shorten the effective horizon, which reduces computation at the cost of potential suboptimality on long-range tasks. Dogar and Srinivasa [54] earlier demonstrated the practical value of combining pushing and grasping in a unified planner for cluttered rearrangement, establishing the push-grasping paradigm that later randomized planners extended and refined.

Human-in-the-loop guidance remains an effective pragmatic response when fully autonomous planning is too slow. Papallas and Dogar [30] demonstrated that injecting minimal operator guidance (a few mouse clicks indicating promising push directions) into a control-based randomized planner substantially reduced planning time and improved success rates, suggesting that the search space contains exploitable structure that pure random sampling fails to leverage. While this human-guided approach is pragmatically valuable, particularly for time-critical teleoperation, it sidesteps rather than solves the underlying combinatorial challenge. The persistent difficulty of fully autonomous randomized planning has driven three broad responses, covered in the remainder of this section. Embedding contact physics directly into continuous optimization (Section 3.2), replacing uniform sampling with learned generative models that focus search on promising regions (Section 3.4), and replacing general-purpose numerical machinery with mode-aware abstractions derived from first-principles contact mechanics (Section 3.5).

3.2 Contact-Implicit Trajectory Optimization

Contact-implicit trajectory optimization (CITO) offers a principled model-based alternative that avoids the combinatorial explosion of explicit contact-mode enumeration by embedding contact constraints directly into a continuous optimization problem. Wang et al. [36] formulated non-prehensile manipulation planning as trajectory optimization with state-triggered constraints, where contact forces are activated and deactivated based on the optimizer's state trajectory rather than pre-assigned to discrete modes. The formulation enables the optimizer to discover contact sequences autonomously, eliminating the need for a priori mode scheduling. Moura et al. [62] pursued a complementarity-constraint variant for non-prehensile planar manipulation, showing that modern mixed-integer and nonlinear solvers can handle contact complementarity directly for practical problem sizes, albeit with longer solve times than smoothed formulations.

Chen et al. [14] extended CITO to three-dimensional non-prehensile transport, demonstrating that incorporating end-effector rotation as an explicit optimization variable, rather than treating the tray orientation as fixed, reduced transport time by exploiting orientation to counteract inertial forces on unstable carried objects. The result establishes that end-effector orientation is a first-class degree of freedom in non-prehensile transport planning, not merely a kinematic convenience. Kitaev et al. [15] pursued a complementary strategy in which physics-simulation rollouts with finite-difference gradients served as the contact model inside trajectory optimization for cluttered grasping. Non-prehensile pushing behaviors (clearing obstacles, avoiding cascade toppling) emerged automatically from the grasp-success objective without any explicit push-planning module, demonstrating that differentiable simulation can serve as a viable alternative to complementarity-based CITO when the contact environment is complex.

3.3 Contact Model Selection and Its Consequences

The choice of contact model within trajectory optimization involves quantifiable trade-offs between physical fidelity, motion quality, and computational cost, and the choice matters more than the quality of the initial guess. Onol et al. [4] conducted a systematic comparison of three contact-model families for manipulation trajectory optimization. Complementarity-based models are mathematically exact but introduce non-smooth constraints, smooth contact models are differentiable but physically approximate, and variable smooth models interpolate between the two. Their analysis revealed that variable smooth models offer the best balance for non-prehensile tasks, and, more importantly, that contact-implicit pushing optimization can succeed from zero-torque initialization, indicating that the contact-model formulation dominates solution quality over initial-guess quality.

Friction uncertainty presents a distinct but related modeling challenge. Zhou et al. [11] addressed it by parameterizing physically consistent friction distributions with a single noise scalar, yielding uncertainty-aware predictions without requiring explicit contact-mode selection. This stochastic formulation complements the smooth-model approach of Onol et al. [4] by adding robustness without adding combinatorial complexity. Chavan-Dafle and Rodriguez [5] demonstrated a different efficiency strategy altogether. By restricting external pushes to sticking contact modes (no pusher-object sliding), they obtained a closed-form dynamics formulation that evaluates 100-1000 times faster than complementarity-based dynamics within sampling-based planners for in-hand manipulation. This contact-mode specialization is orthogonal to both the smooth-versus-complementarity trade-off and the reduced-order abstractions discussed in Section 3.5, offering a targeted acceleration when task physics permit the sticking assumption.

The complementarity constraints (\(0 \leq \boldsymbol{\lambda} \perp \phi(\mathbf{x}) \geq 0\)) that couple contact forces (\(\boldsymbol{\lambda}\)) to the signed gap function (\(\phi\)) are the formal object around which smooth, variable-smooth, sticking-mode, and differentiable-physics relaxations all revolve. Every variant in this family trades a different aspect of the complementarity structure for tractability.

3.4 Decomposing Planning from Control

A recurring architectural insight across the literature is that separating scene-level planning from object-level control yields modular, transferable systems. Mandadi et al. [37] introduced a framework where an A*-based scene-aware planner handles collision avoidance while an object-centric, context-free reinforcement learning controller executes push-by-strike primitives. The controller learns purely from object-state representations and transfers across scenes without retraining. Haustein et al. [29] pursued decomposition at the search level. By replacing uniform random sampling in the planner with a GAN-based generative model that proposes robot configurations from which objects can be transported toward goal states, paired with an RL-learned action policy, they substantially accelerated non-prehensile rearrangement planning by focusing search on promising regions of the composite configuration space.

Wu et al. [16] demonstrated decomposition at the temporal level for long-horizon extrinsic manipulation. By representing a demonstration as an ordered sequence of contact specifications (each primitive's precondition expressed as a state constraint) and retargeting those constraints via inverse kinematics, their system achieved one-shot transfer to novel objects and scenes using a fixed library of short-horizon goal-conditioned primitive policies, attaining approximately 80 percent success across diverse tasks without scene-specific retraining. Lee et al. [63] pushed decomposition into task-and-motion planning for obstacle rearrangement in clutter, interleaving prehensile and non-prehensile primitives within a tree-search framework that exploits the structural differences between the two action types. These four approaches share a common principle, that non-prehensile manipulation decomposes more naturally into layered abstractions than into monolithic end-to-end controllers, but they differ in where the decomposition boundary is drawn. Scene versus object [37], sampling versus dynamics [29], temporal segments versus primitive execution [16], or symbolic task versus continuous motion [63].

3.5 Reactive Hybrid Control and Mode-Aware Reduced-Order Models

An emerging alternative to both full-fidelity contact models and learned surrogates is the construction of compact, mode-aware abstractions from first-principles mechanics. Hogan and Rodriguez [64] first demonstrated real-time reactive planar non-prehensile manipulation by restricting the pusher-slider system to a small library of hybrid modes and solving a mode-scheduled model-predictive control problem at every step. Their approach achieved closed-loop pushing at frame rates unattainable by contact-implicit optimization at the time, at the cost of a hand-engineered mode structure specific to planar pushing. Özcan et al. [7] generalized this mode-aware philosophy. They demonstrated that wrench-twist limit-surface mechanics for planar non-prehensile manipulation can be abstracted into a small discrete library of reduced-order non-holonomic models (unicycle dynamics indexed by contact topology). Pairing this mode-aware abstraction with algebraic force allocation yields an optimization-free, real-time-capable controller for pushing and press-and-slide tasks, eliminating iterative solvers entirely. The approach differs fundamentally from the smooth-model approximations of Onol et al. [4] and the sticking-contact specialization of Chavan-Dafle and Rodriguez [5]. Rather than relaxing or restricting the contact physics, it replaces the full dynamics with a topologically indexed family of simple models, each exact within its contact mode. The result is a system that runs in real time without sacrificing model validity within each mode, a compelling demonstration that domain-specific physical insight can outperform general-purpose numerical machinery for well-structured manipulation problems.

3.6 Multi-Agent and Extended Non-Prehensile Domains

The core computational frameworks for non-prehensile manipulation have been extended to settings that introduce structural complexity beyond single-arm, single-object interaction. Tang et al. [12] formulated collaborative pushing by teams of mobile robots (without manipulators) as hybrid optimization over quasi-statically feasible contact modes shared across multiple agents, with hierarchical path decomposition and per-robot model-predictive-control tracking. Their formulation generalizes the single-arm pushing paradigm by treating under-actuation as arising from constrained multi-contact force distributions across multiple bodies rather than from a single end-effector. Sahin and Chakraborty [13] explored a non-standard actuation modality. Actively modulating contact friction on robot fingers, switching between high- and low-friction surface states, serves as the primary control input for within-hand manipulation, enabling a small primitive library (sliding, rotation, pivoting) to achieve 3D dexterity from a low-degree-of-freedom hand without tactile sensing. This reframes the non-prehensile control problem from "what force to apply" to "what friction to present," opening a design space that complements force-centric planning. Selvaggio et al. [65] approached non-prehensile transport of objects from the model predictive control angle, deriving a non-sliding MPC that guarantees object stability on a moving tray. Rigo et al. [66] extended contact optimization to non-prehensile loco-manipulation with a hierarchical MPC that coordinates whole-body locomotion with contact-rich object interaction. Huang et al. [67] scaled the rearrangement problem to large numbers of objects, revealing that the planning structure that works for two or three objects does not transfer directly to scenes with tens of objects. Lee and Kim [68] extended goal-driven pushing to unknown object properties by treating mass, friction, and inertia as latent variables to be estimated online. Liu et al. [69] demonstrated bimanual non-prehensile manipulation of semi-fluid objects (stir-fry) that couples fluid-like material dynamics with dual-arm coordination, a regime where rigid-body contact models fundamentally break down. Collectively, these extensions show that the principles developed for single-arm planar pushing, contact-mode reasoning, quasi-static models, decomposed planning architectures, transfer to structurally richer domains when the appropriate generalizations of contact mechanics and actuation are identified.

4. Tool Use Reasoning and Affordance Learning

4.1 From Ecological Affordances to Computational Models

The theoretical grounding for computational affordance models in robotics derives from Gibson's ecological psychology, formalized for robotic control by Şahin et al. [46], who proposed a three-part structure. Affordances are relations between agent capabilities, object properties, and action effects. Krüger et al. [45] operationalized this structure through Object-Action Complexes (OACs), grounded abstractions of sensorimotor processes that bind perception, action, and predicted effect into reusable planning units. Montesano et al. [48] provided the first probabilistic implementation, learning affordance distributions from a robot's own sensorimotor interaction with its environment, demonstrating that affordances could be acquired from experience rather than hand-coded, a foundational result that enabled the subsequent proliferation of learning-based affordance models.

Gonçalves et al. [22] extended the self-supervised paradigm to two-object settings, learning visual affordances through autonomous robot exploration where the robot predicts the consequences of its actions on objects without human-provided labels. Jain and Inamura [51] provided a Bayesian formulation that generalizes functional features to estimate the effects of previously unseen tools. Hassanin et al. [70] surveyed the broader visual affordance and function understanding literature, revealing a split between perception-focused affordance segmentation (popular in computer vision) and interaction-focused affordance learning (dominant in robotics). These foundational works established the computational affordance paradigm, which subsequent research has elaborated along three main axes. Richer relational structure (Section 4.3), grasp-conditioned prediction (Section 4.4), and scalable self-supervised discovery (Section 4.5).

4.2 Body Schema Extension and Tool Embodiment

One influential line of work treats tool use not as an external planning problem but as an extension of the robot's internal body representation. Nabeshima et al. [18] demonstrated that a robot could incorporate a tool into its body schema via temporal integration of multisensory signals (vision, touch, proprioception), treating the tool as a functional body extension rather than an external object, enabling reach and interaction planning in body-centric coordinates without explicit geometric tool modeling. The body-schema approach draws on neuroscience evidence that primates neurally integrate tools into their peripersonal-space representations.

Kothavale and Boddepalli [20] pursued a kinematic variant of the body-schema idea. By extending the inverse kinematics solver to treat tool length as an explicit input parameter, a single trained policy generalizes across tools of varying lengths without retraining. This kinematic-layer mechanism for tool affordance adaptation is both simpler and more transparent than perceptual body-schema approaches, although it sacrifices the rich multisensory integration that makes body schema methods robust to visual occlusion. Takahashi et al. [71] bridged the two approaches with a tool-body assimilation model that uses deep learning to integrate grasping motion, tool geometry, and task dynamics into a single latent representation, effectively learning a body schema end-to-end rather than through hand-designed multisensory fusion. Mar et al. [21] showed that grasp-dependent affordance learning on the iCub humanoid implicitly builds a body-tool schema. The robot's functional representation of what it can do with a grasped tool adapts based on how the tool is held, re-parameterizing the body model through the grasp configuration. The body-schema paradigm remains influential because it offers a unified representational framework (the extended body model) that subsumes both tool-tip pose estimation and affordance reasoning, although its reliance on temporal multisensory integration makes it less applicable to one-shot tool encounters.

4.3 Relational Affordance Models for Tool-Target Interaction

The most significant conceptual advance in computational affordance modeling over the review period has been the shift from single-object affordances to relational models that capture the asymmetric structure of tool-target interactions. Gonçalves et al. [22] established the relational framework by modeling affordances as probabilistic dependencies over (action, tool visual features, target visual features, effect) tuples, enabling bidirectional inference (predicting effects from actions or selecting actions to achieve desired effects). Saponaro et al. [24] extended this to cross-modal transfer. Their hand-to-tool affordance model used visual imagination to project tool appearance into a learned hand-posture representation space, enabling zero-shot tool affordance estimation from bare-hand sensorimotor experience without tool-specific training data.

Jamone et al. [25] demonstrated that relational affordance models support multi-step planning through affordance chaining (using the predicted effect of one action as the starting state for the next affordance inference), enabling goal-directed tool-use sequences beyond single-step action selection. Horton and St. Amant [72] developed a partial-contour similarity metric for habile-agent affordance transfer, which matches tools whose silhouettes share functional substructure rather than overall shape. More recently, Tian et al. [52] formalized one-shot 3D object-to-object affordance grounding, arguing that most real-world interactions are inherently relational and that prior work's focus on single-object affordance prediction overlooked this fundamental structure. Ding et al. [27] extended relational modeling to secondary (non-designed) tool affordances, learning from human observation of non-canonical tool actions (a ruler used to push objects) using visual before-and-after object-state changes as the learning signal. The trajectory from single-object to relational to secondary affordance models reflects a progressive expansion of what counts as a relevant affordance feature, from intrinsic tool geometry, to tool-target geometric relations, to use-context-dependent functional properties.

4.4 Grasp-Conditioned and Multi-Granularity Affordance Prediction

A critical empirical finding, replicated across multiple studies, is that tool affordances depend jointly on tool geometry and grasp configuration. The same physical tool affords different actions when grasped differently. Mar et al. [23] demonstrated this through a multi-model approach. Rather than training a single affordance model conditioned on continuous grasp parameters, they first clustered tool-grasp configurations into discrete pose categories (without human labels), then trained a separate affordance model per category, enabling generalization to unseen tools by routing through the nearest discovered category. The discrete multi-model decomposition was motivated by the observation that continuous grasp-conditioned models exhibit poor interpolation in regions of grasp space that are physically discontinuous (power grip versus precision grip, for instance).

Yang et al. [26] advanced this line of work by decomposing affordances into two feature scales. Fine-grained finger-contact-level features spatially localize the functional interaction region on the tool, while coarse-grained hand-object-interaction-level features predict the dexterous grasp gesture. This multi-granularity decomposition treats affordance as a multi-scale phenomenon where different spatial resolutions serve distinct downstream tasks (where to grasp versus how to grasp). Fang et al. [73] addressed the same where-to-grasp question from a task-oriented perspective, learning grasps optimized not for stability but for downstream tool-use effectiveness via simulated self-supervision. Qin et al. [74] introduced KETO, which represents tool function through a small set of semantically meaningful keypoints whose spatial relationships encode functional constraints. Keypoint representations offer a middle ground between dense shape descriptors (expensive to optimize) and parametric models (too rigid for diverse tools), and they transfer across tool instances that share functional structure despite geometric variation.

Zhong et al. [28] proposed a fundamentally different generalization mechanism. A semantic knowledge graph explicitly encodes task-tool-grasp affordance mappings, and a Graph Attention Network propagates relational affordance knowledge from seen to novel task-tool combinations. The graph-based approach offers structured, interpretable generalization rather than interpolation in continuous feature space or clustering-based routing. The diversity of these approaches, discrete clustering, multi-scale feature decomposition, keypoint representations, task-oriented grasp optimization, graph-based propagation, reflects the absence of consensus on the right inductive bias for grasp-conditioned affordance generalization. The question likely depends on the diversity and continuity of the target tool and task distributions.

4.5 Self-Supervised Affordance Discovery

A persistent bottleneck in affordance learning is the acquisition of training labels. Manually annotating which tool-action pairs produce which effects does not scale to the diversity of real-world tool encounters. Mar et al. [21] addressed this through self-supervised affordance category discovery, where a robot autonomously discovers affordance categories by clustering the observed effects of tool-use actions, then uses these emergent categories as teaching signals to train a visual functional-feature-to-affordance classifier. The label-free bootstrapping mechanism was extended by Gonçalves et al. [22], whose autonomous exploration pipeline learned visual affordances from the robot's own interaction history without human supervision.

Tikhanoff et al. [49] demonstrated a related approach on the iCub platform, exploring affordances and tool use through autonomous sensorimotor experimentation. Tee et al. [75] pushed self-supervision toward emergence of tool use without prior tool learning, demonstrating that robots can recognize and use tools through causal observation rather than explicit supervised training on tool categories. Their later work on tool cognition [76] in Nature Machine Intelligence argued that a small set of cognitive primitives (affordance transfer, function abstraction, goal-directed selection) suffices to explain a broad range of animal and human tool behaviors, providing a theoretical framework that has yet to be fully reduced to implementable algorithms. Saponaro et al. [24] achieved a particularly striking form of self-supervised transfer. By projecting tool visual appearance into a hand-posture representation space learned from bare-hand experience, their system estimated tool affordances zero-shot, without any tool-specific interaction data. The common thread across these approaches is that the robot's own sensorimotor experience provides the supervisory signal, with clustering or cross-modal projection replacing human annotation. However, the diversity of affordance categories discoverable through autonomous exploration is bounded by the robot's action repertoire and workspace, a limitation that becomes significant when the target domain includes tools whose affordances require actions outside the robot's exploratory distribution.

4.6 Task-Conditioned Tool Design and Generative Approaches

Rather than selecting strategies for existing tools, a recent line of work asks whether the tool itself can be designed as part of the policy. Liu et al. [17] proposed a task-conditioned tool design policy that learns to generate task-specific tool morphologies via reinforcement learning, treating tool shape as an adaptive affordance variable and enabling zero-shot generalization to unseen goals through joint optimization of design and control policies. The approach inverts the classical affordance problem. Instead of asking what can I do with this tool, it asks what tool should I create for this task, a question that becomes tractable when the tool can be 3D-printed or otherwise fabricated on demand. Ma et al. [31] pursued a different form of generative reasoning through semantic-contact fields, combining semantic planning from vision-language models with precise contact-field representations for tactile tool manipulation, enabling category-level generalization to novel tool instances by grounding high-level functional intent in low-level contact geometry. These generative approaches, whether generating tool shapes or contact-field representations, share the insight that affordance reasoning need not be purely discriminative. Constructing the right intermediary, physical tool or contact representation, can be more effective than classifying affordances of existing intermediaries.

4.7 Online Affordance Refinement and Adaptive Skill Selection

Static affordance models, whether learned from demonstration or self-supervised exploration, are insufficient when object properties change during manipulation. Wu et al. [29] demonstrated this in the domain of robot-assisted feeding. Physical affordance properties of food items (softness, moisture, viscosity) were initially estimated via commonsense VLM inference but required continuous updating through multimodal visuo-haptic perception during ongoing manipulation, as food state changes (steak becoming firm as it cools) invalidate initial estimates. Their SAVOR system treats haptic feedback as a live affordance signal rather than a control-level error correction, enabling mid-task skill re-selection as object properties evolve.

Liu and Asada [50] anticipated this adaptive paradigm in the context of expert tool manipulation. Their process-dynamics model maps task process characteristics (contact dynamics, material state) to control-strategy parameters (feedrate, cutting force, compliance), enabling dynamic, state-adaptive tool strategy switching, treating the process state as the relevant affordance variable rather than static tool or object geometry. Lee and Park [32] explored a related theme in contact tooling for robotic repair, where the manipulation control strategy (hybrid position-force control, admittance control, bilateral telerobotic control) must adapt to the dynamic contact conditions between tool and workpiece. The convergence of these works, from manufacturing skill transfer in 1992 to food manipulation in 2025, establishes that online affordance refinement through closed-loop multimodal sensing is a prerequisite for robust tool use in non-stationary environments.

5. Learning-Based Approaches

5.1 Deep Reinforcement Learning for Non-Prehensile Manipulation

Deep reinforcement learning has emerged as the primary method for acquiring non-prehensile manipulation policies that generalize across diverse geometric and physical environments. Cho et al. [38] demonstrated that modular, reconfigurable network architectures (specialized sub-networks activated based on detected environmental constraints such as walls, steps, and obstacles) enable policies that adaptively switch non-prehensile strategies, achieving generalization across diverse geometries that monolithic policies struggle with. Their hierarchical architecture explicitly addresses the observation that different environmental structures demand qualitatively different manipulation strategies (pushing against a wall versus sliding across an open surface).

Zhou et al. [42] introduced HACMan, which learns hybrid actor-critic maps combining discrete contact-location selection from object point clouds with continuous post-contact motion parameters. The hybrid discrete-continuous action representation substantially outperformed pure discrete or pure continuous baselines in 6D non-prehensile manipulation tasks. Xie et al. [77] earlier foreshadowed this line of work with visual foresight for novel-object tool use. Their system demonstrated that a learned video-prediction model trained on diverse pushing interactions could, at test time, select appropriate "tools" (cans, shoes, blocks) to accomplish a target displacement of another object, rediscovering tool-mediated pushing without explicit tool-use supervision. Raei et al. [44] applied deep deterministic policy gradient specifically to non-prehensile sliding, integrating online friction estimation as closed-loop feedback into the RL actor, enabling real-time adaptation to varying surface friction properties without explicit domain randomization of physics parameters. Collectively, these results establish that the choice of action representation (modular versus monolithic, hybrid versus homogeneous, friction-aware versus friction-agnostic) is at least as consequential as the choice of RL algorithm for non-prehensile manipulation performance.

5.2 Sim-to-Real Transfer for Contact-Rich Manipulation

Zero-shot sim-to-real transfer for non-prehensile manipulation is now achievable through multiple complementary strategies, though the specific combination of techniques varies with the task's contact complexity. Cho et al. [38] demonstrated transfer by combining procedurally generated diverse simulation environments with contact-based geometric representations extended to capture environmental structure, not just object shape, as the observation space. Del Aguila Ferrandis et al. [34] achieved transfer through Bayesian visuotactile state estimators trained via privileged simulation policies, maintaining robust performance under visual occlusions using only a simple onboard camera. Zhou et al. [42] relied on point-cloud-based observations paired with hybrid action spaces, while Raei et al. [44] achieved transfer through online friction estimation that compensates for sim-real friction discrepancies without requiring explicit domain randomization. Park et al. [40] demonstrated that RL-learned non-prehensile skills trained entirely in simulation transfer to semi-autonomous teleoperation contexts where the learned policies serve as suggestion engines for human operators. The diversity of successful transfer strategies, geometric observation augmentation [38], privileged-to-deployable distillation [34], hybrid action spaces [42], online physical-parameter estimation [44], suggests that sim-to-real transfer for non-prehensile manipulation is not a single problem admitting a single solution, but a family of problems where the dominant sim-real gap (visual, geometric, physical, dynamic) determines the appropriate transfer mechanism.

5.3 Multimodal Perception for Robust Non-Prehensile Control

Contact-rich manipulation under partial observability demands perception systems that fuse multiple modalities, and recent work has demonstrated that incorporating tactile sensing into the RL loop yields substantial robustness gains. Del Aguila Ferrandis et al. [34] showed that Bayesian visuotactile state estimators, coupled into the policy's observation space, enable uncertainty-aware non-prehensile manipulation that remains functional under visual occlusions that would defeat vision-only policies. Kim et al. [43] demonstrated that tactile-sensing-based reward shaping, simultaneously incentivizing firm grasping and smooth finger gaiting using contact feedback, improves data efficiency in contact-rich in-hand manipulation and enables sim-to-real transfer to multi-fingered hands without explicit friction modeling. Wu et al. [29] extended multimodal perception to tool-mediated manipulation, showing that visuo-haptic signals enable real-time affordance estimation that adapts to changing object properties during task execution. The progression from vision-only [40] to visuotactile [34, 43] to visuo-haptic with semantic grounding [29] reflects a broader trend toward richer perceptual foundations for contact-rich manipulation, although it also raises questions about the minimum sensor suite required for specific task families, a question that the field has not yet systematically addressed.

5.4 Imitation Learning, Skill Transfer, and Sample Efficiency

Reinforcement learning's requirement for millions of environment interactions has motivated a parallel line of work on imitation learning and structured skill transfer that achieves strong generalization from limited demonstrations. Rana et al. [41] demonstrated that oriented affordance frames, object-centric geometric representations of state and action spaces, enable sample-efficient imitation learning from approximately 10 demonstrations, with compositional generalization to long-horizon, multi-object tasks achieved by composing independently trained sub-policies whose transitions are governed by self-progress prediction derived directly from the timing structure of training demonstrations.

Lu et al. [33] pursued skill transfer across tools rather than scenes. By decomposing tool-use demonstrations into constraint-irrelevant (portable) skill components and separable tool-specific constraint conditions within a Dynamic Movement Primitive (DMP) framework, their method enables cross-tool skill transfer by recombining the portable skill with newly derived constraints, without re-learning from scratch or requiring RL. Zhou and Jia [35] showed that even without RL, bimanual non-prehensile manipulation can be learned from video demonstrations by extracting coarse hand trajectories and refining them with geometry-aware post-optimization, achieving category-level generalization by parameterizing learned skills with object geometric attributes. Nai et al. [53] pushed this paradigm further by eliminating the robot from the demonstration pipeline entirely. Capturing human whole-body motion with portable hardware and translating it through a hierarchical pipeline into feasible humanoid-robot skills achieved 3x faster data collection than teleoperation while enabling tasks that tightly couple locomotion and manipulation. These approaches share a common strategy of imposing geometric or physical structure on the representation to compensate for limited data, but they differ in where that structure is imposed. In the observation space [41], in the skill decomposition [33], in the primitive parameterization [35], or in the motion-retargeting pipeline [53].

5.5 Bimanual, Whole-Body, and Loco-Manipulation

Non-prehensile manipulation research has expanded from single-arm tabletop settings to bimanual, whole-body, and locomotion-coupled manipulation, driven by the recognition that many real-world non-prehensile tasks require coordinated multi-limb interaction. Zhou and Jia [35] introduced BiNoMaP, a framework for learning category-level bimanual non-prehensile manipulation primitives that eliminates dependence on environmental supports (walls, edges) typically exploited by single-arm RL methods, achieving cross-embodiment transfer across kinematically distinct dual-arm platforms without redesigning skill structures. Nai et al. [53] demonstrated humanoid whole-body manipulation, tasks coupling locomotion with manipulation such as kneeling to pick up objects, squatting to push heavy items, and bimanual tossing, learned from robot-free human demonstrations without complex reward engineering or simulation infrastructure. Portela et al. [55] addressed the force-control dimension of loco-manipulation. Their RL policies learn direct force control without force sensing by training with explicit force targets in simulation, enabling compliant whole-body control in legged manipulators where the robot body automatically adjusts posture to achieve force-regulated interactions during teleoperated tool use. The extension to multi-limb and locomotion-coupled settings introduces additional structure (bimanual coordination constraints, balance dynamics, ground-reaction forces) that single-arm methods do not address, but the same decomposition principle, separating skill primitives from their composition and sequencing, appears to scale to these richer embodiments.

5.6 Human-Robot Integration in Learned Manipulation Systems

A pragmatically important but theoretically under-explored mode of deploying learned non-prehensile skills is within human-robot collaborative systems, where learned policies augment rather than replace human decision-making. Park et al. [40] introduced a semi-autonomous teleoperation framework where RL-learned non-prehensile policies serve as suggestion engines, presenting multiple candidate rearrangement actions to a human teleoperator who selects among them, combining machine-generated skill options with human judgment to outperform both fully manual control and fully autonomous execution in cluttered environments. Papallas and Dogar [30] demonstrated the complementary direction. Human guidance injected into autonomous planners, where a few human inputs dramatically improve planning efficiency. Portela et al. [55] showed that RL-learned compliant force control makes teleoperation intuitive by automating postural adjustment while humans command only the manipulator, effectively distributing the control problem between human intent and learned compliance.

Zhu et al. [39] explored VLM-based planning as another form of human-robot interface. Natural language task descriptions are translated into hybrid prehensile and non-prehensile manipulation plans through a digital-twin intermediate layer that enables proactive mental rehearsal and adaptive replanning. These diverse integration modes, human as selector [40], human as guide [30], human as high-level commander [55, 39], suggest that the field is converging toward mixed-initiative systems where the allocation of autonomy between human and robot is itself a design variable, adapted to the task's perceptual and decision complexity. Ravichandar et al. [78] placed these developments in the broader context of learning from demonstration, noting that the field's evolution from low-level trajectory imitation to task-level intent recognition mirrors the trajectory of non-prehensile and tool-use learning specifically.

6. Integrated Tool Use with Non-Prehensile Strategies

6.1 Taxonomic Characterization of Tool Properties for Non-Prehensile Action

A prerequisite for principled tool selection in non-prehensile manipulation is a systematic characterization of which tool properties matter for which actions. Sommer et al. [1] proposed a taxonomy of non-actuated end-effector properties (rigidity, friction coefficient, geometry, including curvature, edge profile, and surface area) that characterizes the affordances of tools for non-prehensile actions such as pressing, rubbing, scraping, and spreading. The taxonomy grounds tool selection in principled physical reasoning rather than ad hoc trial-and-error. Given a target non-prehensile action and its required contact mechanics, the taxonomy identifies which tool-property ranges are compatible. Lee et al. [3] pursued a complementary characterization from the task-planning perspective, using LLM-based symbolic reasoning to decompose tool-object manipulation tasks into sequences of contact actions whose feasibility depends on tool geometry and compliance. The conjunction of bottom-up physical taxonomy [1] and top-down task decomposition [3] provides the beginning of a principled framework for tool selection in non-prehensile contexts, although neither work addresses the problem of uncertain or partially observed tool properties, a gap that online affordance-estimation methods (Section 4.7) are positioned to fill.

6.2 Modular Tool Systems for Extending Gripper Capabilities

The physical integration of tools with robotic end-effectors has been addressed through modular attachment systems that extend a simple gripper's repertoire to non-prehensile actions. Sommer et al. [1] developed a modular tool system for a standard two-fingered parallel-jaw gripper, where task-matched tool modules (pressing pads, rubbing surfaces, scraping edges) can be swapped to enable diverse household and industrial non-prehensile manipulations from a single hardware platform. The approach occupies a pragmatic middle ground between fixed special-purpose end-effectors (limited flexibility) and fully general dexterous hands (excessive complexity for most non-prehensile tasks). Oller et al. [2] addressed the complementary challenge of using an already-grasped tool for non-prehensile manipulation of a target object, a scenario that arises naturally when a robot picks up a spatula, brush, or screwdriver. Their tactile-driven framework explicitly models extrinsic contacts between the grasped tool and the target, using differentiable contact-mode control to optimize robot actions subject to frictional equilibrium constraints. The distinction between tool-as-attachment [1] and tool-as-grasped-object [2] reflects two fundamentally different integration paradigms. In the former, the tool becomes part of the end-effector (rigid attachment), while in the latter, the tool remains a separate object whose coupling to the gripper introduces additional contact uncertainty that must be sensed and controlled.

6.3 Closing the Perception-Action Loop for Tool-Mediated Non-Prehensile Manipulation

The most demanding integration challenge is closing the perception-action loop during tool-mediated non-prehensile manipulation, where the robot must simultaneously estimate the tool-object contact state and adjust the tool's motion in response. Oller et al. [2] demonstrated that tactile sensing coupled with differentiable contact-mode control enables reactive tool manipulation by maintaining explicit models of frictional contact between the grasped tool and the target object, allowing gradient-based optimization of robot actions, a capability that tool-design and tool-attachment frameworks leave as an open interface. Ma et al. [31] proposed semantic-contact fields that combine semantic planning from vision-language models with precise contact representations for tactile tool manipulation, enabling category-level generalization while maintaining the physical grounding required for contact-rich tasks. Lee et al. [3] addressed the planning layer of the loop. Their manoeuvrability-driven controller uses a stepping incremental approach with visual-feedback-derived affordance models to adapt tool trajectories online in spatially confined areas where conventional motion planners fail. The convergence of tactile contact control [2], semantic-contact representation [31], and adaptive trajectory planning [3] suggests that closing the full perception-planning-control loop for tool-mediated non-prehensile manipulation requires advances at all three levels, a systems-integration challenge that no single work has yet fully solved.

7. Cross-Cutting Analysis

7.1 Convergence of Model-Based and Learning-Based Paradigms

The most significant methodological trend across the review period is the erosion of the boundary between model-based planning and control and learning-based approaches. Early non-prehensile manipulation research drew a sharp distinction. Model-based methods (CITO, randomized planning with physics simulators) offered guarantees and interpretability but suffered from computational cost, while learning-based methods (RL, imitation learning) offered speed but lacked guarantees and sample efficiency. By the mid-2020s, the dichotomy has largely dissolved. Haustein et al. [29] injected learned generative models into randomized planners. Mandadi et al. [37] paired model-based A* search with RL controllers. Raei et al. [44] integrated physics-based friction estimation into RL actors. Özcan et al. [7] used first-principles mechanics to derive models fast enough to obviate learning for specific task classes. The most effective systems are not purely model-based or purely learned, but hybrid architectures that use physical models where they are tractable and learning where they are not, with the specific hybrid determined by which aspects of the problem admit closed-form or efficient solutions (quasi-static contacts, sticking-mode dynamics) and which require data-driven approximation (complex contact distributions, high-dimensional perception).

7.2 Representation as the Unifying Challenge

Across all four themes, the choice of state representation, how objects, contacts, tools, and environments are encoded for planning, control, and learning, emerges as the single most consequential design decision. In planning and control, the representation determines computational tractability. Complementarity models [4] versus smooth models versus reduced-order non-holonomic abstractions [7] versus sticking-mode specializations [5]. In affordance learning, the representation determines generalization. Single-object features [49] versus relational tuples [22] versus keypoint abstractions [74] versus knowledge graphs [28] versus multi-granularity decompositions [26]. In RL, the representation determines sample efficiency and transfer. Monolithic observations versus modular environment encodings [38] versus hybrid discrete-continuous action spaces [42] versus oriented affordance frames [41]. In integrated tool-use systems, the representation determines the interface between perception, planning, and control. Physical tool taxonomies [1] versus semantic-contact fields [31] versus contact-mode specifications [2].

No single representation dominates across all settings, suggesting that the field needs either task-adaptive representation selection or a unifying representation framework that spans the fidelity-efficiency spectrum.

7.3 From Isolated Skills to Compositional Systems

A clear temporal trend in the literature is the progression from isolated manipulation primitives to compositional systems that sequence and combine primitives for long-horizon tasks. Early work focused on individual non-prehensile actions (a single push, a single slide) or individual tool-use skills (a single grasp-and-apply). More recent work addresses composition through multiple mechanisms. Contact retargeting of demonstration sequences [16], affordance chaining for multi-step planning [25], VLM-based plan-skeleton generation [39], self-progress prediction for sub-policy sequencing [41], and DMP-based skill decomposition and recombination [33]. The integration of prehensile and non-prehensile skills within a single system, demonstrated by Zhu et al. [39] and Park et al. [40], represents a particularly important compositional advance, as real-world tasks routinely interleave grasping with pushing, sliding, and tool use. However, compositional systems remain fragile. Errors in early primitives propagate and compound through the sequence, and the interfaces between composed skills (what state must primitive A leave the world in for primitive B to succeed?) are typically specified manually rather than learned.

7.4 The Expanding Role of Foundation Models

Vision-language models (VLMs) and large language models (LLMs) have rapidly entered the non-prehensile and tool-use manipulation pipeline as task-level planners and affordance initializers, although their role remains circumscribed to high-level reasoning. Zhu et al. [39] used VLMs to generate plan skeletons that sequence prehensile and non-prehensile primitives, with a digital-twin layer enabling execution monitoring and replanning. Lee et al. [3] used LLMs to decompose natural-language instructions into feasible motion sequences for tool-object manipulation. Wu et al. [29] used VLM inference for initial affordance-property estimation (food softness, moisture) that was subsequently refined through haptic feedback. Ma et al. [31] combined VLM semantic planning with contact-field representations for tactile tool manipulation. The consistent pattern is that foundation models provide coarse semantic and spatial reasoning (selecting which tool to use, which non-prehensile action to attempt, in what order) while physical execution remains governed by model-based controllers or learned policies with explicit contact reasoning. This division of labor is likely to persist. The contact physics of non-prehensile manipulation are sufficiently precise that language-level reasoning cannot substitute for geometric and force-level control, but the task-level sequencing and commonsense reasoning that foundation models provide were previously unavailable from any automated source.

7.5 Sensing Modalities and Their Impact on Manipulation Capability

The review reveals a clear relationship between sensing modality and achievable manipulation complexity. Vision-only systems suffice for pick-and-place in uncluttered environments [40] and for affordance estimation of visible tool properties [22, 21]. Visuotactile systems enable robust non-prehensile manipulation under occlusion [34] and contact-aware in-hand manipulation [43]. Haptic-augmented systems enable adaptive tool use with changing material properties [29, 50]. Force-sensing (or force-estimating) systems enable compliant loco-manipulation [55] and contact tooling [32]. Variable-friction surfaces enable within-hand manipulation without tactile sensing [13]. The progression suggests a rough correspondence. Each additional sensing modality unlocks a class of tasks inaccessible to systems with fewer modalities, while also imposing hardware, calibration, and integration costs. The field has not yet produced systematic guidelines for minimum-viable sensor suites per task class, an important gap for practical deployment.

7.6 Bridging Cognitive Robotics and Control Theory

A final cross-cutting observation is that the most conceptually unified treatments of tool use sit at the intersection of cognitive robotics and control theory rather than inside either field alone. Billard and Kragic [9] argued in a widely cited Science review that the grand challenge of robot manipulation lies precisely in this intersection. The low-level contact physics demand principled control-theoretic treatment, while the high-level task structure demands cognitive-architecture reasoning. The work surveyed in this review confirms that prediction. The most effective integrated systems, semantic-contact fields [31], LLM-guided tool selection [3], relational affordance chaining [25], combine symbolic reasoning drawn from cognitive robotics with contact mechanics drawn from control theory. Qin et al. [57] provided a focused survey of robot tool use from this cognitive-robotics perspective, complementing the largely planning-centric non-prehensile surveys [56] and visual-affordance surveys [70]. The convergence of these traditions has not yet produced a unified theoretical framework, but the empirical pattern across the reviewed works strongly suggests that such a framework is an attainable and useful target.

8. Open Problems and Future Directions

Despite substantial progress, several specific open problems and methodological gaps define the frontier of computational methods for robotic tool use and non-prehensile manipulation.

Unified representations spanning contact fidelity and computational cost. The proliferation of contact representations (complementarity, smooth, variable smooth, reduced-order, sticking-mode, contact fields) reflects the absence of a representation that is simultaneously physically faithful, computationally efficient, and compatible with gradient-based learning. Future work should investigate multi-fidelity contact representations that automatically adjust their complexity based on the current planning or control phase, analogous to adaptive mesh refinement in finite-element analysis. The mode-aware abstractions of Özcan et al. [7] and the variable-smooth models of Onol et al. [4] point toward this direction but remain manually configured rather than automatically adaptive.

Compositional manipulation with formal guarantees. Sequencing non-prehensile and tool-use primitives for long-horizon tasks remains brittle because inter-primitive interfaces are specified informally. Future work should develop compositional verification methods that certify whether the post-condition of one primitive satisfies the pre-condition of the next under contact uncertainty. The contact-retargeting approach of Wu et al. [16] and the affordance chaining of Jamone et al. [25] provide the precondition-postcondition structure on which formal methods could build, but neither work addresses verification under stochastic contact outcomes.

Cross-tool and cross-task affordance transfer at scale. Current affordance models generalize within narrow tool categories (hooks of varying sizes) but struggle with functionally analogous tools of different morphologies (a spatula versus a flat piece of cardboard). The secondary-affordance learning of Ding et al. [27] and the graph-based generalization of Zhong et al. [28] suggest promising directions, but scaling these methods to the diversity of tools encountered in household and industrial settings requires substantially larger and more diverse training distributions than current work employs. Foundation-model-based affordance priors [29, 31] could provide initial broad coverage that is then refined through physical interaction.

Minimum-viable sensing for non-prehensile tool use. The review reveals that different task classes demand different sensor modalities (Section 7.5), but no systematic framework exists for determining the minimum sensor suite required for a given task class and performance level. Future work should develop task-conditioned sensor-selection policies, analogous to the task-conditioned tool-design policy of Liu et al. [17], that recommend the simplest sensor configuration sufficient for a specified manipulation capability.

Real-time adaptation to non-stationary contact environments. Most planning and control methods reviewed assume static or slowly varying contact properties, yet many practical tool-use tasks involve materials that change state during manipulation (food cooking, adhesives curing, surfaces wearing). The online visuo-haptic refinement of Wu et al. [29] and the process-dynamics model of Liu and Asada [50] address this for specific domains, but a general framework for adaptive contact modeling under non-stationary dynamics remains absent. Integrating learned dynamics models with real-time Bayesian updating, extending the stochastic contact model of Zhou et al. [11] to temporal non-stationarity, is a concrete and feasible research direction.

Multi-agent and human-robot non-prehensile collaboration. Tang et al. [12] demonstrated multi-robot collaborative pushing, and Park et al. [40] and Papallas and Dogar [30] demonstrated human-guided non-prehensile manipulation, but no work addresses the combined setting. Multiple robots and humans jointly performing non-prehensile manipulation with shared tools. This combined setting introduces both multi-agent coordination (who pushes what, when) and shared-tool reasoning (who holds the tool, who guides the push) that neither line of work handles in isolation.

Benchmarking and evaluation. Comparing methods across the reviewed themes is hampered by the absence of standardized benchmarks that span the full range of non-prehensile primitives, tool-use tasks, and sim-to-real transfer conditions. Existing manipulation benchmarks tend to emphasize prehensile pick-and-place. A benchmark suite covering planar pushing with varied friction, pivoting and toppling with diverse geometry, tool-mediated scooping and scraping, and semi-fluid manipulation [69] would enable meaningful cross-study comparison and accelerate progress.

9. Conclusion

This scoping review has examined the computational methods, representations, and learning paradigms that enable robotic tool use and non-prehensile manipulation across the period 2014 to 2026. The field has progressed from isolated, single-primitive capabilities toward compositional systems that combine model-based contact reasoning with learned perception and control, although the integration of tool-use reasoning with non-prehensile planning remains nascent, with only a handful of systems [1, 2, 3, 31] addressing the two capabilities jointly. The convergence of model-based and learning-based paradigms, the maturation of relational affordance models, the achievement of zero-shot sim-to-real transfer for contact-rich tasks, and the emergence of foundation models as task-level planners collectively indicate that the technical prerequisites for general-purpose non-prehensile tool use are materializing, albeit not yet unified into a single framework.

The single most important takeaway is that representation, how contacts, affordances, tool-object relations, and manipulation states are encoded, is the dominant factor determining the capability, efficiency, and generalization of manipulation systems across all four themes examined. Advances in algorithms (better optimizers, more sample-efficient RL) and hardware (better tactile sensors, more dexterous hands) yield incremental improvements, while representational innovations (contact-implicit formulations, relational affordance models, hybrid action spaces, semantic-contact fields) yield qualitative capability jumps. The most impactful future research will therefore focus not on improving existing methods within their current representations, but on developing unified, multi-fidelity representations that span the contact physics to semantic-reasoning spectrum while remaining computationally tractable for real-time closed-loop control. The boundary between the robot, the tool, and the environment will ultimately dissolve into a continuous field of contact possibilities in which any object, hand, tool, table surface, or neighboring object, can serve as a mechanical resource.