Embodiment-Based Grouping Strategy

• Embodiment-based grouping strategy is a family of methods that exploit body-related cues such as morphology, joint geometry, and affordance to reduce ambiguous associations.
• It is applied in diverse domains including pose estimation, reinforcement learning, crowd simulation, and scheduling to improve inference, control, and coordination.
• Current research highlights adaptive grouping as a key open challenge, exploring dynamic clustering based on learned embodiment cues and gradient alignment.

to=arxiv_search.search 弘鼎 天天众json {"^^^^^^^^4query4 Offline Reinforcement Learning for Heterogeneous Robot Datasets\" OR ^^^^^^^^4ti:\4 Spatial Context with Graph Neural Network for Multi-Person Pose Grouping\" OR ^^^^^^^^4ti:\4 Crowd Grouping via Heuristic Self-Organization\"","max_results":^^^^4ti:\4query4^^^^,"sort_by":"submittedDate","sort_order":"descending"} RTLU to=arxiv_search.search зхуjson {"^^^^^^^^4query4 OR (&&&^^^^4ti:\4^^^^&&&) OR (&&&^^^^4 OR ti:\4^^^^&&&) OR (&&&^^^^4 OR ti:\4^^^^&&&) OR (Benamara et al., 13 Mar 2026) OR (2405.14073) OR (Xu et al., 2018) OR (Newell et al., 2016) OR (Guckelsberger et al., 2021) OR (Xu et al., 2024) OR (&&&^^^^4ti:\4query4^^^^&&&)","max_results":^^^^4ti:\4 Taken together, recent uses of the phrase “embodiment-based grouping strategy” suggest not a single canonical algorithm, but a family of methods that use embodiment-related structure to decide which entities should be grouped, clustered, or coordinated. Depending on the field, the relevant embodiment signal may be the spatial configuration of human joints, robot morphology, local sensorimotor coupling in crowds, affordance and situatedness in object similarity, or the physical versus virtual presence of interactive agents. The resulting grouping operation is correspondingly diverse: graph partitioning for multi-person pose estimation, static morphology-based clustering for offline reinforcement learning, emergent soft-boundary lanes in crowds, proximity-connected components in trajectory analysis, macro-vehicle formation in traffic scheduling, or agent-composition choices in collaborative human–agent teams (&&&^^^^4query4^^^^&&&, &&&^^^^4ti:\4^^^^&&&, &&&^^^^4 OR ti:\4^^^^&&&, Xu et al., 2024, Benamara et al., 13 Mar 2026).

^^^^4ti:\4^^^^. Conceptual scope and terminology

Across the literature, embodiment denotes different but structurally related constraints. In pose grouping, it is the spatial configuration of body parts and the geometry of relative keypoint offsets. In cross-embodiment robot learning, it is morphology, dynamics, and embodiment graphs built from torso, joints, and feet. In crowd simulation, it is finite-sized discs with bounded speed and local, forward-facing sensing. In object representation, it is affordance and situatedness. In computational creativity, embodiment is organized through a typology comprising structural coupling, historical embodiment, virtual embodiment, physical embodiment, organismoid embodiment, organismic embodiment, and an observed cyborg hybrid (&&&^^^^4query4^^^^&&&, &&&^^^^4ti:\4^^^^&&&, &&&^^^^4 OR ti:\4^^^^&&&, Xu et al., 2024, Guckelsberger et al., 2021).

Domain	Embodiment signal	Grouping outcome
Multi-person pose	Keypoint geometry and appearance	Person-instance partition
Cross-embodiment RL	Morphological similarity	Static robot groups and group-gradient updates
Crowds and trajectories	Local sensing, alignment, proximity	Implicit lanes or persistent components
Traffic merging	Tight headways	Macro-vehicles for scheduling
Object similarity	Affordance and situatedness	Embodied object clusters
Human–agent teams	Physical/virtual embodiment	Team composition and cohesion design

A recurrent source of ambiguity is the proximity of “embodiment-based” to “embedding-based.” “Associative Embedding” is an embedding-based grouping strategy that teaches a single network to output detections and group identities end-to-end, using scalar tags rather than embodiment cues (Newell et al., 2016). This distinction matters because some grouping pipelines compare learned appearance tags, whereas embodiment-based methods explicitly encode morphology, geometry, affordance, or physical presence.

This suggests a useful cross-domain abstraction: embodiment-based grouping exploits constraints induced by bodies, body-like structure, or body–environment coupling to reduce combinatorial ambiguity. The concrete mechanism differs by field, but the recurring role of embodiment is to restrict implausible associations and preserve feasible ones.

^^^^4 OR ti:\4^^^^. Geometry-aware grouping in bottom-up pose estimation

In bottom-up multi-person pose estimation, the grouping problem arises after keypoint detection. “Learning Spatial Context with Graph Neural Network for Multi-Person Pose Grouping” formulates this task as graph partitioning over an undirected fully connected graph PRESERVED_PLACEHOLDER_^^^^4query4^^^^ whose nodes are detections and whose edges preserve the possibility of grouping disjoint visible parts when intermediate joints are missing (&&&^^^^4query4^^^^&&&). Each node carries normalized image coordinates PRESERVED_PLACEHOLDER_^^^^4ti:\4^^^^, a joint type, and a visual descriptor extracted from a PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^ patch and mapped to PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^ dimensions. Each edge carries a geometry embedding derived from the relative offset $d_{mn}=[x_m-x_n,y_m-y_n]$ through a joint-type-conditioned MLP.

The central embodiment branch is a Geometry-aware Association GNN. It learns geometry-based affinity by iteratively aggregating edge embeddings into node representations and then updating pairwise affinities. The node update is

$$z_m^{(l)}=\sum_{k=1}^{N}\tilde{a}^{(l-1)}_{\mathrm{geo},m,k}\cdot edge_{mk},$$

followed by a joint-type-specific MLP, while the edge update regresses a pre-activation score from the two endpoint node features and the edge embedding. The stated intuition is that aggregation pools evidence from all candidate partners across joint types, weighted by the previous affinity estimate, so that a node representation reflects feasible spatial relationships from the whole scene configuration rather than only a local pair. Appearance-based affinity is learned in a separate GNN branch and fused with geometry through an MLP with layer sizes $2$–$16$–$64$–$64$–PRESERVED_PLACEHOLDER_^^^^4ti:\4query4^^^^–PRESERVED_PLACEHOLDER_^^^^4ti:\4ti:\4 after which spectral clustering partitions the learned graph.

The graph-partitioning stage converts the fused affinity matrix to a binary adjacency by thresholding at PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4^^^^, builds the normalized Laplacian PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4^^^^, and minimizes

PRESERVED_PLACEHOLDER_^^^^4ti:\44^^^^

with PRESERVED_PLACEHOLDER_^^^^4ti:\4^^^^5 formed from the eigenvectors associated with the smallest eigenvalues. Clusters are then converted into pose instances, with post-pruning enforcing at most one node per joint type in each selected subset.

Empirically, the method is reported to outperform appearance-only grouping. Geometry-only edge classification accuracy is stated as greater than PRESERVED_PLACEHOLDER_^^^^4ti:\4^^^^6 on COCO graphs. On COCO validation, the Associative Embedding baseline obtains PRESERVED_PLACEHOLDER_^^^^4ti:\4^^^^7 AP, whereas the geometry-plus-appearance method obtains PRESERVED_PLACEHOLDER_^^^^4ti:\4^^^^8 AP; on COCO test-dev, the numbers are PRESERVED_PLACEHOLDER_^^^^4ti:\4^^^^9 AP versus PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4query4^^^^ AP, and with single-person refinement PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4ti:\4^^^^ versus PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4 OR ti:\4^^^^. Using HRNet detector outputs, the comparison is PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4 OR ti:\4^^^^ AP versus PRESERVED_PLACEHOLDER_^^^^4 OR ti:\44^^^^ AP, and with ground-truth keypoints PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^5 AP versus PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^6 AP. On MPII transfer from COCO without retraining, the reported comparison is PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^7 AP versus PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^8 AP (&&&^^^^4query4^^^^&&&).

The contrast with “Associative Embedding” clarifies the specific contribution of embodiment here. Associative Embedding supervises scalar tags so that same-instance detections have similar values and different-instance detections have dissimilar values, but it does not explicitly model global pose geometry (Newell et al., 2016). The geometry-aware GNN instead treats embodiment as scene-level spatial feasibility. A common misconception is therefore that all bottom-up grouping is merely metric learning over appearance tags; in this formulation, grouping is instead grounded in embodied pose topology.

^^^^4 OR ti:\4^^^^. Morphology-grounded grouping in cross-embodiment reinforcement learning

In cross-embodiment offline reinforcement learning, embodiment-based grouping addresses a different failure mode: gradient interference across heterogeneous robot morphologies. “Cross-Embodiment Offline Reinforcement Learning for Heterogeneous Robot Datasets” studies a shared policy PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^9 and shared value functions over robots that differ in dynamics, morphology, and reward coefficients (&&&^^^^4ti:\4^^^^&&&). The paper quantifies inter-robot gradient alignment using pairwise cosine similarity,

PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4query4^^^^

with conflict defined by PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4ti:\4^^^^.

The reported measurements show that the fraction of negative cosine similarities grows both with the proportion of suboptimal data and with the number and diversity of robots. The specific values given are PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4 OR ti:\4^^^^ for Expert, PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4 OR ti:\4^^^^ for PRESERVED_PLACEHOLDER_^^^^4 OR ti:\44^^^^ Suboptimal, and PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^5 for PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^6 Suboptimal; and PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^7 for A^^^^4ti:\4^^^^/Go^^^^4ti:\4 OR ti:\4^^^^, PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^8 for all quadrupeds, and PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^9 for all $d_{mn}=[x_m-x_n,y_m-y_n]$^^^^4query4^^^^ robots. Transfer gain also correlates strongly with mean gradient alignment, with Pearson correlation $d_{mn}=[x_m-x_n,y_m-y_n]$^^^^4ti:\4^^^^ for IQL. The paper therefore clusters robots by morphological similarity before training and updates the actor by group rather than by globally aggregated robot gradients.

Morphology is encoded as a graph $d_{mn}=[x_m-x_n,y_m-y_n]$^^^^4 OR ti:\4^^^^ whose nodes are torso, joints, and feet; whose edges connect torso to adjacent joints, adjacent joints, and terminal joints to feet; and whose node descriptors are the URMA local descriptors used for joints and feet, with the torso represented by a zero vector. Similarity is computed with the Fused Gromov-Wasserstein distance using Euclidean feature ground metric on standardized descriptors, shortest-path structure ground metric, fusion tradeoff $d_{mn}=[x_m-x_n,y_m-y_n]$^^^^4 OR ti:\4^^^^, entropic regularization $d_{mn}=[x_m-x_n,y_m-y_n]$4, and uniform node weights. Distances are min–max normalized to $d_{mn}=[x_m-x_n,y_m-y_n]$5, and hierarchical agglomerative clustering forms $d_{mn}=[x_m-x_n,y_m-y_n]$6 static groups.

Group-wise optimization is actor-only in the recommended configuration. For group $d_{mn}=[x_m-x_n,y_m-y_n]$7, the group loss is

$d_{mn}=[x_m-x_n,y_m-y_n]$8

with gradient

$d_{mn}=[x_m-x_n,y_m-y_n]$9

In practice, the implementation draws a global minibatch, filters the subset for each group, and applies sequential updates over a random permutation of groups within each outer iteration. The critic is updated once per outer iteration from the global minibatch. The authors state that this “update-by-group” design reduces cross-embodiment interference because, within a step, policy updates are shaped by aligned within-group gradients and are not immediately canceled by highly misaligned cross-group gradients.

The empirical gains are substantial in the regime the method targets. Mean performance across six datasets is reported as $$z_m^{(l)}=\sum_{k=1}^{N}\tilde{a}^{(l-1)}_{\mathrm{geo},m,k}\cdot edge_{mk},$$^^^^4query4^^^^ for IQL, $$z_m^{(l)}=\sum_{k=1}^{N}\tilde{a}^{(l-1)}_{\mathrm{geo},m,k}\cdot edge_{mk},$$^^^^4ti:\4^^^^ for IQL+PCGrad, $$z_m^{(l)}=\sum_{k=1}^{N}\tilde{a}^{(l-1)}_{\mathrm{geo},m,k}\cdot edge_{mk},$$^^^^4 OR ti:\4^^^^ for IQL+SEL, and $$z_m^{(l)}=\sum_{k=1}^{N}\tilde{a}^{(l-1)}_{\mathrm{geo},m,k}\cdot edge_{mk},$$^^^^4 OR ti:\4^^^^ for IQL+EG. On $$z_m^{(l)}=\sum_{k=1}^{N}\tilde{a}^{(l-1)}_{\mathrm{geo},m,k}\cdot edge_{mk},$$4 Suboptimal Forward, IQL improves from $$z_m^{(l)}=\sum_{k=1}^{N}\tilde{a}^{(l-1)}_{\mathrm{geo},m,k}\cdot edge_{mk},$$5 to $$z_m^{(l)}=\sum_{k=1}^{N}\tilde{a}^{(l-1)}_{\mathrm{geo},m,k}\cdot edge_{mk},$$6 with EG; on $$z_m^{(l)}=\sum_{k=1}^{N}\tilde{a}^{(l-1)}_{\mathrm{geo},m,k}\cdot edge_{mk},$$7 Suboptimal Backward, from $$z_m^{(l)}=\sum_{k=1}^{N}\tilde{a}^{(l-1)}_{\mathrm{geo},m,k}\cdot edge_{mk},$$8 to $$z_m^{(l)}=\sum_{k=1}^{N}\tilde{a}^{(l-1)}_{\mathrm{geo},m,k}\cdot edge_{mk},$$9. The average improvement on the $2$^^^^4query4^^^^ splits is stated as $2$^^^^4ti:\4^^^^ relative to the IQL baseline, compared with $2$^^^^4 OR ti:\4^^^^ for SEL and $2$^^^^4 OR ti:\4^^^^ for PCGrad. The ablation on grouping strategy reports $2$4 for the IQL baseline, $2$5 for random groups, $2$6 for a heuristic split, and $2$7 for embodiment groups on $2$8 Suboptimal Forward (&&&^^^^4ti:\4^^^^&&&).

Related unsupervised cross-embodiment pre-training provides a complementary perspective. “PEAC: Unsupervised Pre-training for Cross-Embodiment Reinforcement Learning” formalizes a Controlled Embodiment MDP and learns an embodiment discriminator $2$9 with intrinsic reward $16$^^^^4query4^^^^ (2405.14073). A plausible implication is that embodiment grouping need not rely exclusively on explicit morphology graphs; discriminator embeddings, occupancy signatures, or skill signatures can also serve as clustering substrates when the objective is to organize embodiments by behavior or dynamics similarity rather than by morphology alone.

4. Emergent, topological, and scheduling-based grouping in embodied motion systems

In motion systems, embodiment-based grouping often replaces learned affinity with explicit local interaction rules or topological connectivity. “Emergent Crowd Grouping via Heuristic Self-Organization” models each agent as a finite-sized disc with bounded speed and local, forward-facing sensing (&&&^^^^4 OR ti:\4^^^^&&&). Grouping is implicit rather than explicit: agents rotate their preferred velocity according to neighbors that are close, visible, and moving in similar directions, and detour away from neighbors likely to cause head-on interactions. The preferred-velocity rotation is written as

$16$^^^^4ti:\4^^^^

with $16$^^^^4 OR ti:\4^^^^, implying $16$^^^^4 OR ti:\4^^^^. The rotated preference is then supplied to ORCA, which solves

$16$4

The paper reports that, across Circle, AsyCircle, ^^^^4 OR ti:\4^^^^-Group, and 4-Group scenarios, the method achieves the lowest congestion in some classic scenarios and yields lower, more stable discrepancy after the initial transient than Social Forces, RVO, PLEdestrians, PowerLaw, Karamouzas, Moussaïd, ORCA, Dutra, and Implicit Crowds.

“Trajectory Grouping Structures” treats grouping as continuous-time connectivity over trajectories $16$5 (&&&^^^^4ti:\4query4^^^^&&&). At each time $16$6, it defines a proximity graph

$16$7

and a group $16$8 exists when $16$9 stays inside a proximity-connected component throughout a time interval $64$^^^^4query4^^^^, with $64$^^^^4ti:\4^^^^ and $64$^^^^4 OR ti:\4^^^^. The full evolution of groups, merges, splits, births, and deaths is encoded by a Reeb graph. The paper proves that the Reeb graph has worst-case size $64$^^^^4 OR ti:\4^^^^, that the number of maximal groups is $64$4, and that robust maximal groups under $64$5-relaxed connectivity can be computed in $64$6 time. Here, embodiment enters through the geometric trajectories and the interaction radius $64$7; grouping is topological rather than learned.

“A Grouping Based Cooperative Driving Strategy for CAVs Merging Problems” applies grouping to scheduling at a single-lane mainline and single-lane on-ramp merge (Xu et al., 2018). Consecutive vehicles on the same lane whose initial time headway satisfies $64$8 are absorbed into a single macro-vehicle, so that only the interleaving of groups across lanes remains to be optimized. This reduces the search space from

$64$9

To cap computation, the procedure starts from $64$^^^^4query4^^^^ and increases $64$^^^^4ti:\4^^^^ by $64$^^^^4 OR ti:\4^^^^ until the number of groups is at most $64$^^^^4 OR ti:\4^^^^. Runtime at $64$4 veh/(lane·s) is reported as $64$5 s for planning, $64$6 ms for grouping, and $64$7 ms for ad hoc negotiation; at $64$8, as $64$9 s, PRESERVED_PLACEHOLDER_^^^^4ti:\4query4query4^^^^ ms, and PRESERVED_PLACEHOLDER_^^^^4ti:\4query4ti:\4^^^^ ms. Average delay comparisons include PRESERVED_PLACEHOLDER_^^^^4ti:\4query4 OR ti:\4^^^^ s/veh for grouping versus PRESERVED_PLACEHOLDER_^^^^4ti:\4query4 OR ti:\4^^^^ for ad hoc at PRESERVED_PLACEHOLDER_^^^^4ti:\4query44^^^^, and PRESERVED_PLACEHOLDER_^^^^4ti:\4query4^^^^5 versus PRESERVED_PLACEHOLDER_^^^^4ti:\4query4^^^^6 at PRESERVED_PLACEHOLDER_^^^^4ti:\4query4^^^^7. In this setting, embodiment is encoded by kinematics, lane membership, and safety headways rather than by morphology or social embodiment.

These examples show that “grouping” can be implicit, explicit, or structural. In crowds it is a soft-boundary self-organization effect. In trajectory analysis it is a maximal connected component over time. In CAV merging it is a combinatorial reduction device. What unifies them is the use of embodied constraints to eliminate infeasible or low-value interactions before or during coordination.

5. Social embodiment, perceived credibility, and collaborative team composition

In interactive systems, embodiment-based grouping is not limited to clustering observations; it also concerns how agents are composed into teams and how embodiment changes perceived group quality. “To Embody or Not: The Effect Of Embodiment On User Perception Of LLM-based Conversational Agents” studies embodied and non-embodied conversational agents in cooperative, non-hierarchical survival tasks (&&&^^^^4 OR ti:\4^^^^&&&). The embodied condition uses a high-fidelity MetaHumans avatar in Unreal Engine with VITS speech, NVIDIA Omniverse Audio^^^^4 OR ti:\4^^^^Face, and real-time streaming responses from LLaMA ^^^^4 OR ti:\4^^^^.^^^^4ti:\4^^^^ 8B Instruct. The non-embodied condition is text-only, with the same LLM.

Contrary to the common expectation that embodiment improves social evaluation, the reported quantitative results favor the non-embodied system on competence: PRESERVED_PLACEHOLDER_^^^^4ti:\4query4^^^^8 versus PRESERVED_PLACEHOLDER_^^^^4ti:\4query4^^^^9 on a PRESERVED_PLACEHOLDER_^^^^4ti:\4ti:\4query4^^^^–PRESERVED_PLACEHOLDER_^^^^4ti:\4ti:\4ti:\4^^^^ scale, with Wilcoxon PRESERVED_PLACEHOLDER_^^^^4ti:\4ti:\4 OR ti:\4^^^^. Character, sociability, and dynamism also trend in favor of the non-embodied condition, with sociability at PRESERVED_PLACEHOLDER_^^^^4ti:\4ti:\4 OR ti:\4^^^^ versus PRESERVED_PLACEHOLDER_^^^^4ti:\4ti:\44^^^^ and PRESERVED_PLACEHOLDER_^^^^4ti:\4ti:\4^^^^5. Mean sentiment of user messages is significantly more positive toward the non-embodied agent, with PRESERVED_PLACEHOLDER_^^^^4ti:\4ti:\4^^^^6. Qualitatively, PRESERVED_PLACEHOLDER_^^^^4ti:\4ti:\4^^^^7 participants explicitly criticized the embodied CA for insufficient pushback, versus PRESERVED_PLACEHOLDER_^^^^4ti:\4ti:\4^^^^8 for the non-embodied CA. The paper theorizes that embodiment increases social expectations for authenticity, so sycophantic agreement without justification becomes more damaging to perceived competence and credibility. In this literature, embodiment-based grouping therefore includes deployment rules such as non-embodied lead agents for high-consequence factual decision-making and mixed configurations pairing a non-embodied “competence anchor” with a more restricted embodied partner.

“Exploring the role of embodiment on intimacy perception in a multiparty collaborative task” shifts the problem from dyadic CA perception to multiparty human–agent teams (Benamara et al., 13 Mar 2026). It presents a corpus and a pre-registered protocol for groups of two humans and two agents under three conditions: two ECAs, two Furhat robots, or one of each. The Furhat robot head provides physical co-location, projected face, lip movement, built-in speaker, and head motion; the ECA condition uses ^^^^4 OR ti:\4^^^^D animated agents on a vertically oriented display, with facial expressions, lip movements, and upper-body gestures. The planned measures include the Virtual Intimacy Scale, QAG/GEQ for task and social cohesion, ESAS for belonging and acceptance, SPRS and SSC for social and collaborative skills, sociogram choices, task logs, gaze, emotion, gestures, and speech.

No inferential statistical results on embodiment effects are reported in that paper, and the authors explicitly describe the current corpus as small and unbalanced (Benamara et al., 13 Mar 2026). Accordingly, the proposed grouping strategy is forward-looking: hybrid Robot–ECA compositions are presented as a way to combine physical salience with the “safer” space for disclosure associated with virtual agents; two identical-looking agents are discouraged because they risk being perceived as a bonded subgroup; and agent roles and turn-taking are to be standardized through a shared finite-state machine so that embodiment, rather than role assignment, drives any observed difference. This is a distinct sense of grouping: not clustering data instances, but composing a socially coherent mixed group.

6. Embodied similarity spaces and typologies of embodied systems

A further usage of embodiment-based grouping appears in perceptual representation. “Object Space is Embodied” argues that perceived object similarity is shaped not only by visual or conceptual features but also by embodied features: affordance and situatedness (Xu et al., 2024). Affordance is defined as relatively stable action possibility, such as power grasp, precision grasp, virtual fingers, or posture closeness. Situatedness is the transient spatial relation to an agent, including palm orientation, finger orientation, movement direction, wrist twist, elbow posture, and egocentric distance. The paper operationalizes these notions through action atoms and deterministic annotation rules, then constructs embodied representational dissimilarity matrices.

For coffee mugs, embodied features explain subjective similarity with adjusted PRESERVED_PLACEHOLDER_^^^^4ti:\4ti:\4^^^^9, affordance-only features with adjusted PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4query4^^^^, and situatedness-only features with adjusted PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4ti:\4^^^^, while visual-only AlexNet features explain adjusted PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4 OR ti:\4^^^^. The unique embodied contribution after controlling for visual features is reported as PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4 OR ti:\4^^^^. For novel objects, embodied features explain adjusted PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\44^^^^, affordance-only features PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4^^^^5, situatedness-only features PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4^^^^6, and visual-only features PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4^^^^7. The subjective mug space is described as PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4^^^^8-dimensional with non-metric MDS stress below PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4^^^^9; embodied feature space has effective dimensionality PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4query4^^^^, whereas visual feature space has effective dimensionality PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4ti:\4^^^^ (Xu et al., 2024). In this setting, grouping means clustering objects in a feature space whose axes are explicitly embodied.

“Embodiment and Computational Creativity” provides a more general taxonomy for grouping systems by embodiment profile rather than by task behavior (Guckelsberger et al., 2021). It distinguishes structural coupling, historical embodiment, virtual embodiment, physical embodiment, organismoid embodiment, organismic embodiment, and an observed cyborg embodiment. The review reports that, across ^^^^4 OR ti:\4query4ti:\4query4^^^^–^^^^4 OR ti:\4query4 OR ti:\4query4^^^^ ICCC papers, explicit, non-metaphorical embodiment appears in PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4 OR ti:\4^^^^ papers; structural and physical embodiments are most common, at approximately PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4 OR ti:\4^^^^ each; historical, virtual, and organismoid appear in approximately PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\44^^^^ papers; and organismic appears in PRESERVED_PLACEHOLDER_^^^^4ti:\4 OR ti:\4^^^^5, typically human-focused. The proposed feature-vector view, with dimensions for structural coupling, history, physical or virtual substrate, organismoid similarity, organismic autonomy, and cyborg involvement, offers a general-purpose grouping framework for embodied creative systems.

These two literatures push embodiment-based grouping beyond coordination and inference. They treat embodiment as a representational axis along which objects or systems can be organized, compared, and clustered. A plausible implication is that grouping by embodiment can function either as an optimization aid, as in RL and traffic, or as a descriptive model of similarity structure, as in object space and computational creativity.

7. Limitations, misconceptions, and open directions

A first misconception is that embodiment is uniformly beneficial. The conversational-agent study reports the opposite in its setting: the non-embodied agent was rated significantly more competent, and the embodied agent attracted more sycophancy complaints (&&&^^^^4 OR ti:\4^^^^&&&). A second misconception is that embodiment-based grouping must be dynamic. In cross-embodiment offline RL, the reported best-performing method uses static morphology-based clusters computed once before training, although the same paper notes that fixed clusters may become suboptimal as learning progresses (&&&^^^^4ti:\4^^^^&&&). A third misconception is that embodiment-based grouping and embedding-based grouping are interchangeable; the distinction between geometry-aware embodiment and associative embedding remains substantive (Newell et al., 2016).

The limitations are also domain-specific. In pose grouping, incomplete detections under heavy occlusion can still confuse the embodiment branch, leaving clusters that require appearance to disambiguate (&&&^^^^4query4^^^^&&&). In crowd self-organization, highly heterogeneous goals, abrupt goal changes, and tight bottlenecks can cause chattering or residual stop-and-go waves (&&&^^^^4 OR ti:\4^^^^&&&). In multiparty embodiment studies, the current evidence on cohesion and intimacy is still methodological rather than inferential, because the reported corpus is underpowered and unbalanced (Benamara et al., 13 Mar 2026). In object-space work, the reported action atoms were developed for coffee mugs and a specific set of novel shapes, which suggests that broader object categories may require expanded embodied descriptors (Xu et al., 2024).

The major open direction across fields is adaptive grouping. The RL literature explicitly lists dynamic or online grouping based on morphology and live gradient statistics as a possible extension (&&&^^^^4ti:\4^^^^&&&). The CE-MDP and embodiment-discriminator framework in PEAC suggests an additional route in which grouping is updated from learned embodiment posteriors rather than from a fixed morphology graph (2405.14073). In social systems, the open problem is not only how to group agents, but how to align embodiment choice with task demands such as intimacy, critique, and psychological safety (Benamara et al., 13 Mar 2026, &&&^^^^4 OR ti:\4^^^^&&&).

In this broader sense, embodiment-based grouping is best understood as a methodological principle rather than a single algorithm: group entities according to the structure imposed by bodies, morphology, affordance, physical presence, or sensorimotor coupling, and use that structure either to improve inference and control or to describe the organization of complex interactive systems.