Geometric Amoebot Model for Programmable Matter

• Geometric Amoebot Model is a lattice-based specialization with particles occupying triangular (or FCC) grids, enabling inherent geometric properties like boundary and directional cues.
• It supports distributed algorithms for leader election, shape formation, and reconfiguration under diverse execution models including asynchronous, synchronous, and circuit-augmented paradigms.
• Recent extensions, such as 3D variants, joint movements, and reconfigurable circuits, accelerate global tasks like symmetry breaking and geometric decomposition, broadening programmable matter capabilities.

Searching arXiv for recent and foundational papers on the geometric amoebot model and its extensions. The geometric amoebot model is the lattice-based geometric specialization of the canonical amoebot model for programmable matter: anonymous particles occupy nodes of a regular lattice, communicate through locally labeled ports, and reconfigure by expansions and contractions while maintaining a connected occupied structure (Daymude et al., 2015). In its standard two-dimensional form, the ambient space is the infinite regular triangular grid $G_\Delta$, but the literature now includes a three-dimensional FCC variant, a reconfigurable-circuit extension, and a joint-movement extension; together these developments have turned the model into a common framework for studying distributed symmetry breaking, shape formation, routing, recognition, locomotion, and global geometric decomposition in programmable matter (Briones et al., 2022, Padalkin et al., 2022, Padalkin et al., 2023, Hillebrandt et al., 17 Apr 2026).

1. Canonical geometric formulation

The foundational distinction is between the general amoebot model and the geometric amoebot model. In the general model, particles inhabit an arbitrary graph $G=(V,E)$; in the geometric model, the underlying graph is fixed to the equilateral triangular graph, equivalently the triangular lattice $G_\Delta$, which is dual to the hexagonal tiling frequently used in the informal exposition of programmable matter (Daymude et al., 2015). This specialization supplies geometric notions absent from the general model, such as boundary, clockwise order, line, polygonal turning angle, and directional adjacency, and those notions are exactly what early leader-election and shape-formation algorithms exploit (Daymude et al., 2015).

Particles are homogeneous computational elements. In the standard geometric model, a particle may be contracted, occupying one node, or expanded, occupying two adjacent nodes; at most one particle occupies a node at any time, and motion is realized by expansion and contraction, with handovers available in the more general formalism (Daymude et al., 2015, Alumbaugh et al., 2019). The occupied nodes induce a connected subgraph, and connectivity preservation is a standing correctness requirement across essentially all reconfiguration work in the area (Daymude et al., 2015). In stationary submodels, particles remain contracted throughout execution and the algorithmic problem is solved purely by communication and state change rather than movement, as in the three-dimensional leader-election work and much of the circuit literature (Briones et al., 2022, Padalkin et al., 2022).

Local geometry is encoded through ports and orientation. Foundational two-dimensional work assumes that each particle labels incident bonds consecutively in clockwise order and that all particles share chirality but not necessarily a global compass, so local port numbers are comparable only up to rotation (Daymude et al., 2015). Later work weakened that assumption: deterministic universal shape formation was shown without chirality by explicitly establishing a common handedness procedurally, and other reconfiguration work even treats particles as disoriented, with no shared global orientation and no common chirality (Luna et al., 2017, Kostitsyna et al., 2022). This divergence is a recurrent theme: the geometric amoebot model is best understood not as a single fixed axiom set, but as a geometric core together with a family of assumption variants.

2. Locality, symmetry, and execution models

The model’s algorithmic behavior is controlled as much by symmetry and scheduling as by lattice geometry. A decisive concept is unbreakable rotational symmetry. Deterministic shape-formation work on simply connected initial configurations defines a shape to be unbreakably $k$-symmetric if it has a center of $k$-fold rotational symmetry that does not coincide with a grid vertex; on the triangular grid only $k\in\{2,3\}$ can occur (Luna et al., 2017). The central impossibility theorem states that if the initial shape is unbreakably $k$-symmetric, then any deterministically formable minimal target shape must also be unbreakably $k$-symmetric (Luna et al., 2017). In that sense, symmetry rather than chirality is the fundamental obstruction to deterministic formation.

Execution models vary sharply across the literature. The early geometric-model papers and several reconfiguration papers operate under asynchronous or adversarial activation: the 2015 leader-election and line-formation paper uses an asynchronous adversarial scheduler with one active particle at a time; the 2017 universal shape-formation result works under a strong adversarial scheduler not necessarily sequential; direct reconfiguration via shortest-path transport is proved under a sequential scheduler and then extended to the asynchronous setting; and the Tile Automata simulation paper linearizes asynchronous amoebot executions into a fair sequential scheduler for analysis (Daymude et al., 2015, Luna et al., 2017, Kostitsyna et al., 2022, Alumbaugh et al., 2019). By contrast, most circuit-enabled and joint-movement results assume fully synchronous rounds, because those extensions rely on highly coordinated global signaling or movement phases (Feldmann et al., 2021, Padalkin et al., 2022, Padalkin et al., 2024, Padalkin et al., 2023, Artmann et al., 28 Jan 2025, Kumar et al., 11 Mar 2026, Hillebrandt et al., 17 Apr 2026).

Topological assumptions are equally important. Some algorithms require connected and contractible structures, as in the FCC-lattice leader-election algorithm in three dimensions (Briones et al., 2022). Some require hole-free triangular-grid configurations, as in the shortest-path-forest algorithms based on portal trees (Padalkin et al., 2024). Other papers are expressly about overcoming those restrictions: the 2026 geodesically convex decomposition paper treats arbitrary connected structures with holes and proves a decomposition into $\Theta(|\mathcal H|)$ simple geodesically convex regions in $O(\log n)$ rounds w.h.p. (Hillebrandt et al., 17 Apr 2026). Thus “the” geometric amoebot model supports a broad spectrum of algorithmic regimes, from local asynchronous motion on arbitrary simply connected shapes to globally synchronized, circuit-augmented computations on topologically unrestricted inputs.

3. Core two-dimensional algorithmics

The early 2D literature established that geometry fundamentally changes what anonymous constant-memory particles can do. The 2015 paper “Leader Election and Shape Formation with Self-Organizing Programmable Matter” separates the general amoebot model from the geometric one and proves that leader election and line formation become solvable in the geometric setting precisely because the triangular lattice supplies boundary cycles, clockwise traversal, and polygonal angle information (Daymude et al., 2015). In that paper, leader election is decided by a boundary-cycle algorithm with expected $G=(V,E)$0 rounds, while line formation from a unique leader is achieved with worst-case optimal $G=(V,E)$1 work (Daymude et al., 2015). The same paper also proves impossibility of deciding leader election and line formation in the general, geometry-free amoebot model, making geometry a model-theoretic resource rather than a mere embedding convenience (Daymude et al., 2015).

The 2017 universal shape-formation paper considerably extends this picture. Starting from any simply connected initial configuration of $G=(V,E)$2 particles, it gives a deterministic universal algorithm that forms any target shape of constant base size compatible with the symmetry class of the initial configuration, without assuming chirality and under a strong adversarial scheduler (Luna et al., 2017). Its complexity is $G=(V,E)$3 rounds and $G=(V,E)$4 moves, and the move complexity is asymptotically optimal because any universal shape-formation algorithm requires $G=(V,E)$5 moves in the worst case (Luna et al., 2017). The same work shows that if randomization is permitted, the remaining deterministic symmetry obstruction disappears: provided $G=(V,E)$6 is large enough, any target shape can be formed from any simply connected initial shape (Luna et al., 2017).

A different extension of expressive power comes from the 2020 “Mobile RAM” construction. That paper shows that four amoebot particles can simulate a mobile two-register RAM, hence a mobile Turing-complete machine, by encoding unbounded memory geometrically in inter-particle distances rather than in per-particle state (Luna et al., 2020). It then uses that construction to obtain a universal shape-formation algorithm for computable shape families, including circles, spirals, and finite-stage fractal objects such as the Sierpinski triangle and the Koch snowflake (Luna et al., 2020). This result substantially broadens the geometric amoebot model’s expressive scope: shape formation is no longer limited to line and triangle assemblages, but can be compiled from arbitrary Turing-computable lattice descriptions.

Reconfiguration work later shifted from canonical intermediate shapes to geometry-sensitive transport. The 2022 paper “Fast Reconfiguration for Programmable Matter” routes only the symmetric difference $G=(V,E)$7 between initial and target shapes through the persistent core $G=(V,E)$8, rather than dismantling the entire structure into a line or triangle first (Kostitsyna et al., 2022). Its key primitives are shortest-path maps, cone-grown shortest-path trees, feather trees, and coarse-grid junction control; the resulting algorithm solves the reconfiguration problem in $G=(V,E)$9 rounds under a sequential scheduler, $G_\Delta$0 rounds in expectation asynchronously, and $G_\Delta$1 rounds with high probability in the asynchronous setting, with an $G_\Delta$2 lower bound showing worst-case optimality (Kostitsyna et al., 2022). This strand illustrates a broader shift in the field: once geometric routing structures become available, the model supports transport and reconfiguration algorithms that preserve large portions of the occupied set rather than rebuild from scratch.

4. Reconfigurable circuits and global geometric computation

The most consequential complexity-theoretic extension of the geometric amoebot model is the reconfigurable circuit model. In this extension, each adjacency between neighboring amoebots is replaced by a constant number of external links, each endpoint being a pin. Every amoebot partitions its pins into disjoint partition sets; over the full system, these partition sets form a graph whose connected components are circuits (Padalkin et al., 2022). A beep sent on a partition set in one round is received at the beginning of the next round by all partition sets in the same circuit, but only as a Boolean event: receivers learn neither sender identity nor multiplicity (Padalkin et al., 2022). This is a weak communication primitive in content, but a strong one in topology, because the effective communication diameter of a circuit is one.

The immediate algorithmic consequence is acceleration of symmetry-breaking and agreement tasks that are diameter-bound in the local-only model. “Accelerating Amoebots via Reconfigurable Circuits” proves leader election in $G_\Delta$3 rounds w.h.p., consensus in $G_\Delta$4 rounds, compass alignment in $G_\Delta$5 rounds w.h.p., chirality agreement in $G_\Delta$6 rounds w.h.p., parallelogram recognition with linear or polynomial side ratio in $G_\Delta$7 rounds w.h.p., and recognition of connected minimal shapes composed of triangles in $G_\Delta$8 rounds if common chirality is given, or $G_\Delta$9 rounds otherwise (Feldmann et al., 2021). The same paper also proves that chirality agreement without movement is impossible with only one pin per bond, even for two adjacent particles, so the power of circuits depends on nontrivial hardware assumptions (Feldmann et al., 2021).

The 2022 paper “The Structural Power of Reconfigurable Circuits in the Amoebot Model” generalizes this acceleration phenomenon from symmetry breaking to structural geometry (Padalkin et al., 2022). It introduces a reusable methodology based on PASC, chain and stripe identifiers, and canonical skeletons. Within this framework, stripe computation is solved in $k$0 rounds; directional global maxima in $k$1 rounds w.h.p.; canonical skeleton and canonical skeleton path construction in $k$2 rounds w.h.p.; spanning-tree computation in $k$3 rounds w.h.p.; and symmetry detection in $k$4 rounds w.h.p. (Padalkin et al., 2022). These results are significant because they recast global geometric inference as circuit-enabled arithmetic on dynamically defined chains and boundary sets.

Routing tasks benefit similarly. The 2024 shortest-path-forest paper works on hole-free amoebot structures and exploits portal graphs in the three lattice directions, together with the identity

$k$5

to reduce shortest-path computation to rooted tree processing on implicit portal trees (Padalkin et al., 2024). It computes a shortest-path tree for one source and $k$6 destinations in $k$7 rounds, which implies $k$8-round single-pair shortest paths and $k$9-round single-source shortest paths, and it computes an $k$0-shortest path forest for $k$1 sources in $k$2 rounds (Padalkin et al., 2024). The hole-free assumption is essential, because the portal graphs cease to be trees in the presence of holes (Padalkin et al., 2024).

Containment and decomposition extend the circuit agenda from routing to global shape analysis. The 2025 shape-containment paper studies the problem of finding maximally scaled embedded copies of a target shape without moving particles (Artmann et al., 28 Jan 2025). It proves a lower bound of $k$3 synchronous rounds for the general problem, then introduces the classes of snowflake shapes and star convex shapes; for star convex shapes the maximum scale can be computed in $k$4 rounds, while for snowflake shapes that are not star convex the bound is $k$5 (Artmann et al., 28 Jan 2025). The structural reason is exact: the paper proves that self-containedness is equivalent to star convexity, so binary search on scale is possible precisely for star convex targets (Artmann et al., 28 Jan 2025).

The most powerful current circuit-enabled geometric result is the 2026 decomposition algorithm for arbitrary structures with holes (Hillebrandt et al., 17 Apr 2026). In the reconfigurable circuit extension of the geometric amoebot model, it computes a decomposition of any connected amoebot structure into $k$6 simple geodesically convex regions in $k$7 rounds w.h.p., where $k$8 is the number of inner holes (Hillebrandt et al., 17 Apr 2026). The construction proceeds in three phases: cutting along selected portals to obtain simple regions, refining them into tunnel regions, and then using additional directional portal cuts to obtain convex subregions, with a delicate final treatment of a possible middle region $k$9 via median portals $k\in\{2,3\}$0 (Hillebrandt et al., 17 Apr 2026). The same paper also improves the global-maxima algorithm of Padalkin et al. for boundary sets to $k\in\{2,3\}$1 rounds w.h.p., which in turn improves their spanning-tree algorithm to $k\in\{2,3\}$2 rounds w.h.p. in the corresponding special case (Hillebrandt et al., 17 Apr 2026). This result is especially revealing about the model: once global geometry is lifted into dynamically reconfigured circuits, even shortest-path-sensitive decomposition tasks on shapes with holes become polylogarithmic-round computations.

5. Three-dimensional variants, joint movements, and simulations

The geometric amoebot model has also been generalized beyond its standard 2D local-motion form. The most direct spatial extension is the 3D geometric space variant formalized on the face-centered cubic lattice $k\in\{2,3\}$3 (Briones et al., 2022). In that model, each occupied FCC node lies in exactly four triangular lattices, and an amoebot’s local orientation is given by view, spin, and rotation, for a total of $k\in\{2,3\}$4 possible orientations (Briones et al., 2022). Using an erosion-based algorithm that preserves connectedness and contractibility, the paper deterministically elects exactly one leader in $k\in\{2,3\}$5 rounds, in fact within $k\in\{2,3\}$6 rounds, under an unfair sequential adversary for connected contractible 2D or 3D systems (Briones et al., 2022). By applying a concurrency-control transformation, it also obtains the first known amoebot leader-election algorithm correct under an unfair asynchronous adversary in both 2D and 3D, though no asynchronous runtime bound is given (Briones et al., 2022).

A different extension changes motion rather than ambient space. The 2023 paper “Reconfiguration and Locomotion with Joint Movements in the Amoebot Model” formalizes the joint-movement extension, in which synchronized bond release followed by expansion, contraction, and handover allows particles to push and pull attached substructures (Padalkin et al., 2023). On top of that semantics it builds rhombical and hexagonal meta-modules that support slide, rotation, tunneling, and related modular-robot primitives, yielding centralized shape-formation algorithms for $k\in\{2,3\}$7 rhombical meta-modules in $k\in\{2,3\}$8 rounds and for $k\in\{2,3\}$9 hexagonal meta-modules in $k$0 rounds, as well as locomotion constructions based on rolling, crawling, and walking (Padalkin et al., 2023). The point of this extension is not only faster reconfiguration but a qualitatively different motion model: coordinated motion of large rigid-looking substructures becomes native.

The 2026 paper “Sublinear-Time Reconfiguration of Programmable Matter with Joint Movements” sharpens the algorithmic implications of that extension (Kumar et al., 11 Mar 2026). Under centralized control, it proves that any connected geometric amoebot structure can be reconfigured into a canonical line segment in $k$1 rounds, thereby giving the first within-the-model sublinear-time universal reconfiguration result without auxiliary assumptions such as metamodules (Kumar et al., 11 Mar 2026). It also shows that any spiral structure can be transformed into a line segment in $k$2 rounds (Kumar et al., 11 Mar 2026). These results rely on constant-time primitives such as tunneling on alternating chains, shearing, parallelogram transfer, triangle reconfiguration, and trapezoid reconfiguration, and they show that joint movements substantially change the reconfiguration complexity landscape.

Finally, the model has been embedded into other formal systems. The 2019 Tile Automata paper gives a constructive simulation of arbitrary geometric amoebot systems by active tile-based self-assembly at scale factor $k$3 (Alumbaugh et al., 2019). In that simulation, specially shaped macrotiles tessellate the square grid so that their adjacency graph is isomorphic to the triangular lattice, preserving contracted and expanded occupancy, local port semantics, handovers, and local communication through flags (Alumbaugh et al., 2019). This does not alter the amoebot model itself, but it demonstrates that its geometric dynamics can be compiled into a self-assembly formalism with local attachment, detachment, and state transitions.

6. Scope, misconceptions, and current frontier

Several common misunderstandings disappear once the literature is viewed as a whole. First, the geometric amoebot model is not identical to “programmable matter on a triangular grid” in a single fixed technical sense. The foundational 2D model, the 3D FCC variant, the reconfigurable-circuit extension, and the joint-movement extension all preserve the same geometric-particle ethos, but they support very different algorithmic possibilities (Daymude et al., 2015, Briones et al., 2022, Padalkin et al., 2022, Padalkin et al., 2023). Second, the strongest runtimes in the literature usually rely on explicit assumptions beyond the local asynchronous base model: fully synchronous rounds, common compass orientation, common chirality or a preprocessing phase that establishes it, hole-free or contractible inputs, reconfigurable circuits, or centralized scheduling (Padalkin et al., 2024, Artmann et al., 28 Jan 2025, Briones et al., 2022, Kumar et al., 11 Mar 2026). Those assumptions do not invalidate the results, but they locate them within a specific part of the model family.

Third, the model should not be conflated with broader bio-inspired particle systems. The 2012 Physarum-inspired multi-agent model is highly relevant as an analog of local sensing, occupancy exclusion, indirect coordination, and coherent shape change, but it is not a formal geometric amoebot model: it uses a discrete 2D lattice with field-mediated stigmergy and single-step particle translations rather than expansion/contraction semantics, handovers, or port-based local protocols (Jones et al., 2012). It is best understood as a biologically inspired cousin rather than as an amoebot paper in the formal programmable-matter sense (Jones et al., 2012).

The current frontier is therefore not a single open problem but a set of interacting ones. Deterministic shape formation is understood up to symmetry constraints in the local 2D model (Luna et al., 2017). Circuit-augmented models now support polylogarithmic-time algorithms for recognition, routing, containment, and decomposition, yet shortest-path problems with holes and broader containment classes remain open (Padalkin et al., 2024, Artmann et al., 28 Jan 2025, Hillebrandt et al., 17 Apr 2026). Joint movements permit sublinear centralized reconfiguration, but distributed fast reconfiguration under local communication remains unresolved, and the literature explicitly points to reconfigurable circuits as a plausible ingredient for overcoming the standard $k$4 barrier there as well (Kumar et al., 11 Mar 2026). A plausible implication is that the most important contemporary question is no longer whether the geometric amoebot model can support global geometric computation, but which combinations of geometry, communication, and motion assumptions are minimally sufficient for each class of global tasks.