Primitive Imbuing: Cross-Domain Insights

• Primitive imbuing is the systematic process of endowing systems with basic operational structures, observed in topology, robotics, graphics, algebra, and VM engineering.
• It employs methods like Heegaard-surface embeddings, state-dependent motion primitives, and geometric assemblies to parameterize and enhance system behavior.
• The approach transforms unstructured operations into tractable substrates, enabling efficient learning, semantic assembly, and dynamic runtime generation across multiple fields.

“Primitive Imbuing” is not a standardized term shared across the cited literatures. As an Editor’s term, it denotes a recurring operation in which a mathematical object, robot policy, 3D model, assembly plan, or runtime system is endowed with primitive structure, or is shown to possess primitiveness only through a specific embedding, factorization, or execution substrate. The arXiv record uses “primitive” in several technically distinct senses: as a Heegaard-surface property of knot embeddings, as a reusable motion primitive in contact-rich manipulation, as a geometric primitive in 3D abstraction, as a semantic carrier in craft assembly, as a notion in permutation groups and words, and as a low-level VM primitive generated on demand (Doleshal, 2011, Wu et al., 2022, Ye et al., 7 May 2025, Araújo et al., 2015, Chari et al., 2013).

1. Conceptual scope and cross-domain uses

A common misconception is that “primitive” names a single cross-domain property. The literature does not support that reading. In topology, primitiveness is positional: it depends on how a knot sits on a genus-2 Heegaard surface. In robot learning, primitives are predefined state-dependent skills whose parameters are adapted from demonstrations. In graphics, primitives are typed geometric parts such as cuboids or ellipsoids. In assembly planning, primitive solids serve as inventory elements onto which semantic and functional roles are assigned. In algebra and combinatorics, primitiveness concerns either the action of a group or the non-power status of a word. In VM engineering, primitives are low-level operations whose implementations may themselves be generated at runtime.

Domain	Primitive object	Sense of imbuing
Knot theory	$K\subset F$ on a genus-2 Heegaard surface	primitiveness is attached to a specific embedding
Robot learning	state-dependent motion primitives	demonstrations and BO tune reusable insertion skills
3D abstraction	cuboids, elliptical cylinders, ellipsoids	shapes are converted into editable primitive assemblies
Craft assembly	cuboids and cylinders from inventory	semantic parts are instantiated as primitive solids
Group and word theory	primitive groups; primitive words	primitiveness is tested under synchronization, concatenation, or insertion
VM systems	VM primitives and plug-ins	primitives are generated, replaced, or customized at runtime

This suggests that “primitive imbuing” is best understood as a family of structurings rather than a single theory. The common thread is that primitive structure is treated as operative: it constrains what can be proved, learned, edited, validated, or executed.

2. Heegaard-surface embeddings and twisted torus knots

In knot theory, the most literal form of primitive imbuing appears in the study of knots lying on the standard genus-2 Heegaard surface $F$ of $S^3=H\cup_F H'$. A knot $K\subset F$ is primitive with respect to $H$ if $H[K]$ is a solid torus, and Seifert with respect to $H$ if $H[K]$ is a Seifert fibered space; likewise for $H'$. Hence $K$ is primitive/primitive if both $F$0 and $F$1 are solid tori, and primitive/Seifert if one side is a solid torus and the other is Seifert fibered. The associated surface slope on $F$2 is the isotopy class of a component of $F$3. In this formulation, primitiveness is not merely a property of the knot type but of the specific embedding $F$4 (Doleshal, 2011).

The same paper studies twisted torus knots $F$5, obtained from the torus knot $F$6 by twisting $F$7 strands $F$8 full twists, and realizes them naturally on the standard genus-2 Heegaard surface. For $F$9 and $S^3=H\cup_F H'$0, the knot $S^3=H\cup_F H'$1 is represented as the closure of

$S^3=H\cup_F H'$2

in $S^3=H\cup_F H'$3. The central fiberedness result is Theorem 3.1: $S^3=H\cup_F H'$4 The proof rewrites the braid as a positive braid word. For $S^3=H\cup_F H'$5, the explicit representative

$S^3=H\cup_F H'$6

is obtained; for general $S^3=H\cup_F H'$7, positivity follows from

$S^3=H\cup_F H'$8

Because positive braid closures are fibered by Stallings’ theorem, the inequality $S^3=H\cup_F H'$9 converts negative twisting into fiberedness.

The same argument delineates the limits of the phenomenon. The paper gives $K\subset F$0, represented by

$K\subset F$1

as a negatively twisted torus knot that is isotopic to the 5-crossing twist knot and is not fibered. Thus primitive/Seifert behavior does not force fiberedness in general.

The interaction with primitive/Seifert embeddings is particularly direct for Dean’s family

$K\subset F$2

When $K\subset F$3, these knots fall under Theorem 3.1 and are fibered. The paper’s second major theorem, Theorem 4.1, then shows that

$K\subset F$4

with

$K\subset F$5

are isotopic as knots in $K\subset F$6 and have the same surface slope

$K\subset F$7

but no homeomorphism of $K\subset F$8 sends $K\subset F$9 to $H$0. For $H$1, this yields a p/p-p/S family; for $H$2, a p/S-p/S family. The result is central to any primitive-imbuing reading of the paper, because it demonstrates that primitiveness and Seifertness are genuinely positional properties of the Heegaard-surface placement.

3. Primitive-based sensorimotor skill acquisition

In robot learning, primitive imbuing denotes endowing a robot with a reusable library of low-level insertion skills and adapting them by parameter search rather than learning a monolithic end-to-end policy. “Prim-LAfD” proposes exactly this architecture for contact-rich insertion. The setting comprises eight real-world tasks: six 3D-printed peg-hole geometries—round, triangle, parallelogram, rectangle, hexadecagon, ellipse—plus RJ45 and waterproof connector tasks. Experiments inject uncertainty of $H$3 mm in translation and $H$4 in orientation along each dimension, so recovery from misalignment must be performed through compliant contact behavior (Wu et al., 2022).

The policy is implemented as state-dependent motion primitives executed by a state machine on top of a Cartesian impedance controller. The state is

$H$5

and the action consists of a desired Cartesian pose $H$6 and diagonal stiffness values $H$7. Each primitive $H$8 specifies a trajectory generator $H$9, an exit condition $H[K]$0, and a stiffness vector $H[K]$1, with all task parameters collected in

$H[K]$2

The four primitives are structurally fixed:

• Free space alignment: moves to an initial alignment pose above the hole, with $H[K]$3.
• Move until contact: descends until force threshold $H[K]$4 is exceeded, with $H[K]$5.
• Search: performs compliant lateral and yaw search using a Lissajous pattern with fixed $H[K]$6, with $H[K]$7.
• Insertion: executes the final compliant push, with $H[K]$8.

The learning signal does not clone demonstrations directly. Instead, 10 demonstrations per insertion task, each averaging 43.5 time steps, are modeled as state trajectories. A rollout $H[K]$9 induced by primitive parameters is assigned likelihood

$H$0

with $H$1 estimated by a Gaussian Mixture Model using $H$2 clusters. The optimization target is

$H$3

where $H$4 is a sparse bonus for task success. This dense objective is the key imbuing mechanism: demonstrations define what successful primitive executions should look like, while Bayesian Optimization tunes the primitive parameters.

Training uses 40 iterations of BO per task; each iteration executes the current policy twice with independently sampled perturbed hole poses, and the average objective is returned to BO. The best policy is evaluated over 20 independent trials. Adaptation to unseen tasks transfers a reduced search box from the top $H$5 similar tasks, where similarity is the $H$6 distance between turning functions of hole cross-section contours. The transferred search region is

$H$7

with

$H$8

The experimental platform is a 6-DoF FANUC Mate 200iD equipped with an ATI Mini45 force/torque sensor; policy outputs are consumed at 10 Hz, while torque commands are streamed at 1000 Hz. The method learns skills in about 40 iterations, taking less than one hour, and adapts to unseen tasks in as few as 15 iterations, or less than 15 minutes on average. Across the eight tasks, LfD outperforms Time on almost all tasks; on parallelogram and RJ45, the Time baseline finds no success within the 40-iteration budget, whereas LfD finds successful parameters for all tasks. In this literature, primitive imbuing is therefore a structured alternative to end-to-end policy learning: the primitive library is fixed, physically meaningful, and reused across tasks, while learning acts on parameters and transfer ranges.

4. Learned 3D primitive assemblies and human-aligned abstraction

In 3D vision and graphics, primitive imbuing takes the form of assigning a human-interpretable primitive assembly to an arbitrary shape. “PrimitiveAnything” reformulates shape primitive abstraction as primitive assembly generation rather than per-shape geometric fitting. The target representation is an ordered set

$H$9

where each primitive has a class label $H[K]$0, scale $H[K]$1, rotation $H[K]$2, and translation $H[K]$3. The supported primitive families are cuboids, elliptical cylinders, and ellipsoids, all expressed in a unified transform interface (Ye et al., 7 May 2025).

The training corpus is HumanPrim, with 120K samples, each containing a 3D mesh, its surface point cloud, and a manual primitive assembly. Primitive sequences average 30.9 primitives, with a maximum of 144; family frequencies are 85.2% cuboids, 11.8% elliptical cylinders, and 3.0% ellipsoids. Data are normalized to a unit cube, encoded as point clouds, and conditioned by the Michelangelo encoder. Transform attributes are discretized rather than regressed continuously: rotations use 180 levels per dimension, scales 128, and translations 128.

A major technical issue is parameter ambiguity under symmetry. The paper defines a symmetry-aware canonicalization using the rotational symmetry set

$H[K]$4

and selects the equivalent parameterization with minimal rotation $H[K]$5 norm. This produces a unique target encoding for symmetric primitives such as cuboids.

The generation model is a shape-conditioned decoder-only transformer with three learned components: primitive encoder $H[K]$6, decoder-only transformer $H[K]$7, and cascaded primitive decoder $H[K]$8. Primitive sequences are ordered by centroid in z-y-x order, and an explicit $H[K]$9 decoder permits variable-length output. The cascaded decoder predicts

$H'$0

in the order “type $H'$1 place $H'$2 orient $H'$3 size”. Training uses

$H'$4

with Gumbel-Softmax for differentiable sampling and Chamfer distance between sampled primitive point clouds.

Quantitatively, on the HumanPrim test set against EMS and Marching-Primitives, PrimitiveAnything reports CD 0.0404, EMD 0.0475, Hausdorff 0.158, and Voxel-IoU 0.484. On segmentation alignment it reports RI 0.892, VOI 2.296, and SC 0.409. On 20 Objaverse shapes rated by 30 participants, it achieves 4.17 geometric similarity, 4.18 anthropomorphism, and 4.22 editability. Ablations show that removing ambiguity-free parameterization degrades CD from 0.0404 to 0.0564 and Voxel-IoU from 0.484 to 0.414. The representation also reduces storage by over 95% compared to mesh representations. In this setting, primitive imbuing is not merely decomposition; it is a learned, human-aligned construction process yielding typed, editable, variable-length primitive programs.

5. Functional craft assemblies over primitive inventories

“Prompt2Craft” is only partially a primitive-abstraction paper. Its main contribution is an LLM-centered craft assembly system rather than a new primitive-fitting method. Here primitives matter chiefly in two places: the available inventory consists of primitive solids, and a baseline inherits an older primitive simplification step from prior work. The Craft Assembly Task takes an RGB image of a target object in the wild, a textual description of available objects, a target function label—one of "hit", "support", or "rolling"—and a structured category template. The output is a valid structured file specifying components, selected available objects, orientations, optional modifications, connectivity relationships, and which parts execute the target function (Isume et al., 4 Dec 2025).

The available inventory contains 41 object types, all composed of cuboids and cylinders, with unlimited quantity. Assembly geometry is constrained by orthogonal positional relations, predefined orientation conventions, specific connection types, and a limited modification type "HOLE". Generated assemblies must pass format validation, collision validation, connectedness validation, and physics-based functional validation. For the proposed method, there is no explicit visual part segmentation stage, no template mesh retrieval stage, and no mesh pose optimization stage. Primitive simplification appears only in the baseline, where manually segmented images via SAM2 are passed to PartCrafter, and the resulting parts are simplified to primitive shapes using prior work.

The paper is explicit that, for the proposed method, primitive assignment is semantic and symbolic rather than analytic. Candidate objects are primitive solids with known dimensions ranging from 10 mm to 250 mm along principal axes. The LLM infers semantic parts, chooses available objects, determines poses, proposes modifications, and assigns functional roles. In that sense, a primitive is “imbued” with part identity, orientation, placement relation, connectivity type, and an exec_function flag.

Evaluation uses 800 images total, 100 per category, with OpenAI o4-mini-2025-04-16 as the LLM and NVIDIA Isaac Sim at 500 Hz for physics validation. Functional success reaches 85.0% under Re-prompt, compared with 63.375% w/o Re-prompt; + Collision feedback gives 75.75%, and + Collision + Simulation feedback 77.125%. The main cause of failure is collision validation, accounting for 69.1\% of the errors. Visual similarity, measured against TripoSG proxy meshes, yields average Chamfer Distance 0.401, Hausdorff Distance 0.630, and [email protected] 0.325 for the proposed method, versus 0.395, 0.630, and 0.360 for PartCrafter*. The paper therefore supports a narrower notion of primitive imbuing than PrimitiveAnything: primitives serve as a robust interface for semantic assembly reasoning, not as the output of a new primitive-fitting objective.

6. Algebraic and combinatorial notions of primitiveness

In permutation-group theory, primitive imbuing has a different meaning. A group $H'$5 acting on a finite set $H'$6 is synchronizing if, for every non-permutation $H'$7, the semigroup $H'$8 contains a constant map. A synchronizing group is necessarily primitive, but the converse fails. “Primitive groups and synchronization” disproves the conjecture that primitive groups synchronize every non-uniform transformation by exhibiting primitive groups that fail to synchronize specific non-uniform transformations of ranks 5 and 6. It also proves that a primitive group of degree $H'$9 synchronizes every transformation of rank $K$0 or $K$1. In this literature, synchronization is a transformation-theoretic strengthening of primitivity rather than a consequence of it (Araújo et al., 2015).

In combinatorics on words, a nonempty word is primitive if it is not a proper power. “Combinatorial Properties of primitive words with Non-primitive Product” gives a complete characterization of distinct primitive words $K$2 such that

$K$3

The three mutually exclusive forms are: $K$4 or

$K$5

or

$K$6

The paper then counts such exceptional pairs via

$K$7

Here primitiveness is not positional but combinatorial, and the question is when concatenation destroys it (Echi et al., 2022).

A related but distinct refinement is insertion robustness. “Ins-Robust Primitive Words” defines $K$8 as the set of primitive words that remain primitive after insertion of any single letter at any position. The complement $K$9 is characterized by

$F$00

Equivalently, a primitive word is non-ins-robust iff some cyclic permutation is periodic with period $F$01 satisfying

$F$02

The paper proves that $F$03 is dense, that $F$04 is not context-free, and that ins-robustness can be recognized in linear time by detecting suitable periodic factors in $F$05 (Srivastava et al., 2017). This suggests a broad combinatorial reading of primitive imbuing: primitiveness may be stable, fragile, or recoverable depending on the operation—action, concatenation, or insertion—that is applied.

7. Runtime generation and modification of VM primitives

In VM engineering, primitive imbuing becomes literal runtime acquisition of low-level capabilities. “Waterfall: Primitives Generation on the Fly” presents Waterfall, a dynamic and reflective translator from Slang to native code, implemented at language-side in the Pharo/Cog setting. Waterfall parses Slang into an AST, optionally lowers through a chain of converters, emits native code, resolves VM symbols through Benzo, and installs the generated code so that primitives and plug-ins can be altered while the VM is already running. The paper’s stated contribution is to allow altering and configuring components such as primitives and plug-ins at runtime; Waterfall “generates primitives on demand and executes them on the fly” (Chari et al., 2013).

The architecture centers on ClosureNativizer, a hierarchy of SendNativizers, a set of Primitives used as pre-nativized templates for operations such as arithmetic and direct memory access, and chained Converters. Primitive methods are identified by a special pragma; Waterfall removes the pragma, installs code that triggers on-demand generation, and then reuses the generated native entry point on subsequent calls. It currently does not support polymorphism, only modularization, and complete symbol-resolution functionality is available only on Unix platforms through dlsym and parsing of nm.

A central case study modifies the essential primitive basicNew. The Waterfall wrapper $F$06 adds instrumentation by calling printOop before delegating to primitiveNew. This is primitive imbuing in the strongest systems sense: a running VM is dynamically endowed with a new low-level behavior around object allocation.

Performance remains below that of the static primitive, but the dynamic approach is substantially faster than reflective instrumentation. For object creation, the paper reports 0.28 ± 0.16 ms for Unmodified, 21.80 ± 0.33 ms for Unsafe reflective instrumentation, 27.72 ± 0.40 ms for Secure reflective instrumentation, 7.72 ± 0.27 ms for Waterfall-based instrumentation, and 7.08 ± 0.23 ms for Waterfall-based unmodified. The file plug-in case study further shows that dynamically generated plug-in functionality can be almost identical in performance to the static plug-in for directory creation. In this literature, primitive imbuing is not metaphorical: it is runtime generation, replacement, and customization of the primitive layer itself.

Across these literatures, primitive imbuing consistently marks a move from unstructured or monolithic description toward a primitive substrate that is operationally consequential. In topology, the substrate is the Heegaard-surface embedding; in robotics, a library of motion/contact primitives; in graphics, a variable-length assembly of typed geometric parts; in craft planning, an inventory of primitive solids carrying semantic and functional roles; in algebra and word theory, a structural condition tested under composition; and in VM systems, the executable primitive interface. This suggests that the deepest commonality is not the meaning of “primitive” but the methodological role of primitive structure as the level at which behavior, classification, editability, and control become tractable.