Distributed Swarm Morphogenesis with Neural Cellular Automata

I. INTRODUCTION

• The paper considers local-relative sensing without globally referenced robot poses, motivated by limits of GPS, motion capture, or centralized estimation in infrastructure-poor settings.The introduction identifies GNSS-denied environments as one such setting.Source: §I · p. 1
• The framework separates candidate-site discovery from assignment and control, allowing existing assignment and control methods to coordinate robots to occupy discovered sites.The paper uses artificial-light-field assignment and constrained motion control from.Source: §I · p. 1
• Offline sparse-retention fine-tuning simulates partial attachment to address the mismatch between virtual NCA growth and gradual physical assembly.A false-positive loss also penalizes site activations outside the desired morphology.Source: §I · p. 1
• Robots exchange and update local NCA states to identify candidate attachment sites, enabling progressive formation without a complete site list.Local NCA computations generate global structure without a central coordinator.Source: §I · p. 1

II. BACKGROUND AND RELATED WORK

• Prior morphology representations include continuous regions, density fields, discrete lattices, and graphs; this work adopts square lattices for well-defined attachment sites.The representation encodes spatial relationships.Source: §II · p. 1
• Progressive self-assembly uses decentralized local rules, while gene-regulatory models, reaction–diffusion dynamics, and learned NCAs provide other development approaches.The cited self-assembly rules are described as rigid.Source: §II · pp. 1–2
• Earlier NCA robotics applications include simulated soft-robot regeneration and physical modular-robot shape classification, but not learned site discovery combined with progressive assembly by independently mobile robots.The paper states, to the authors' knowledge, that this combination had not previously been studied.Source: §II · p. 2
• The NCA uses neighborhood message passing on a regular lattice, with a shared symbolic frame preserving neighbor directions instead of a general-graph GNN formulation.The paper therefore formulates progressive site generation in the simpler NCA setting.Source: §II · p. 2
• Assembly methods range from task allocation and robot-to-site assignment to assignment-free movement, while localization and control use global positioning or relative observations.The implementation uses local relative observations and reuses the constrained controller from.Source: §II · p. 2

III. METHODOLOGY

• Offline training teaches an NCA to reconstruct desired morphologies and fine-tunes it for gradual assembly; deployment keeps its weights fixed while robots update local states to discover sites.The method comprises offline training and deployment phases.Source: §III · p. 2

A. Problem Setting

• N homogeneous holonomic robots assemble using local sensing and communication; an attached Advertiser updates NCA states and advertises sites, while an unattached Joiner seeks them.A designated seed remains stationary, and an attached Joiner becomes an Advertiser.Source: A. Problem Setting · p. 2
• The desired morphology is a finite connected set of square-lattice cells containing the seed, but its complete geometry is available only during offline training.At most one robot occupies each cell, and the morphology contains no more cells than robots.Source: A. Problem Setting · p. 2

B. Local Reference Frame Estimation

• A Joiner estimates a planar rigid transform from observed Advertiser positions and their symbolic cells to locate candidate sites in its body frame.With only the seed as reference, orientation is undetermined; the first Joiner establishes a lattice axis, and later estimates require two attached robots with distinct known coordinates to be observable.Source: B. Local Reference Frame Estimation · p. 2

III. METHODOLOGY

• Advertisers maintain local NCA patches and refresh them from observed attached neighbors before advertising adjacent unoccupied foreground cells.Joiners can occupy advertised cells and then become Advertisers, progressively growing the formation.Source: Figure 1 · p. 3

• Each Advertiser maintains a persistent 5 × 5 state patch and applies the shared NCA rule using each cell’s 3 × 3 neighborhood.The patch supplies context for updating its occupied center and adjacent candidate cells.Source: C. Candidate-Site Discovery · p. 3
• At each control step, Advertisers refresh their patches with nearby occupied-cell states received from other Advertisers before updating the NCA.Source: C. Candidate-Site Discovery · p. 3
• An Advertiser proposes an unoccupied cardinal neighbor only when its updated alpha value exceeds the foreground threshold, sharing the site coordinate and state.Source: C. Candidate-Site Discovery · p. 3
• A Joiner selects an advertised unoccupied cell and, after attaching, initializes its state from an advertised vector and can advertise further sites.Source: C. Candidate-Site Discovery · p. 3

D. Site Selection and Attachment

• The ALF assignment favors advertised sites with lower local light intensity, while the constrained controller balances target tracking, collision avoidance, and velocity smoothing.Both components are replaceable within the framework.Source: D. Site Selection and Attachment · p. 3

E. NCA Training and Fine-Tuning

• The NCA is trained offline to reconstruct each desired morphology, then fine-tuned with sparse retention to account for gradual robot assembly.Source: E. NCA Training and Fine-Tuning · p. 3

1) Growing NCA Training:

• Growing NCA training uses neighborhood features, a shared network’s predicted state change, and stochastic per-cell update masks before optimizing final-state reconstruction.Training unrolls H steps from a living seed and uses backpropagation through time.Source: 1) Growing NCA Training · pp. 3–4

• Sparse-retention rollouts simulate gradual assembly by updating virtual state, retaining selected foreground candidates adjacent to the occupied morphology, and pruning unsupported state.The process continues from the newly occupied frontier.Source: 2) Sparse-Retention Fine-Tuning · p. 4
• The retention distribution can select no candidates, all candidates, or a random subset to model slow and intermittent arrivals.Source: 2) Sparse-Retention Fine-Tuning · p. 4
• Virtual NCA states are kept only on occupied cells and their immediate neighbors, and ordinary rollouts are mixed in to preserve original developmental dynamics.Source: 2) Sparse-Retention Fine-Tuning · p. 4
• A false-positive penalty targets above-threshold foreground activation in background cells and is combined across the rollout with final RGBA reconstruction loss.Source: 2) Sparse-Retention Fine-Tuning · p. 4
• Fine-tuning encourages local physical growth while suppressing false-positive sites; the complete target morphology and training losses are used only offline.Source: 2) Sparse-Retention Fine-Tuning · p. 4

• The evaluation includes simulations of formation progression and scaling, a fine-tuning ablation, and physical quadrotor experiments.The passage describes the study plan but does not report outcomes.Source: §IV · p. 4
• Morphology accuracy is measured by F1 overlap between expected and occupied cells, with exact matches scoring 1 and missing or incorrectly occupied cells lowering the score.Source: §IV · p. 4
• Saturation time is the earliest time from which F1 remains within 5% of its terminal value.Source: §IV · p. 4

Formation and Scaling

• The formation study evaluates Lizard, Tree, Pretzel, and Hand-heart with 250 robots, and varies robot count for Tree.Figure 2 shows representative rollouts and scaling results.Source: A. Simulation Results · pp. 4–5, Figure 2 · p. 5
• Rollouts form branched, compact, and holed morphologies without a deployment-time goal list, with final F1 values close to 1.Tree forms faster than narrow-path or constrained-branch Lizard and Hand-heart; saturation time grows approximately linearly with robot count while final F1 decreases slowly.Source: A. Simulation Results · p. 5, Figure 2 · p. 5

Ablation Study

• Sparse-retention fine-tuning increases final F1 relative to standard Growing NCA by training on partial assembly states.The ablation isolates attachment-site generation from assignment and control.Source: A. Simulation Results · p. 5
• Adding false-positive loss after sparse-retention further improves mean F1 and significantly reduces variation across runs.The ablation evaluates deployment F1 at saturation for Tree morphology with 250 robots.Source: A. Simulation Results · p. 5, Table II · p. 5

Hardware Validation

• Lighthouse tracking and centralized computation reconstruct local-frame observations because the platform lacks the required relative-position sensing.The resulting observations feed the framework, which returns body-frame velocity commands while preserving its local information structure.Source: B. Hardware Validation · p. 5
• Physical experiments show progressive attachment-site discovery and robot recruitment forming the desired E and 4 morphologies.The execution is demonstrated despite localization noise, actuation error, and asynchronous motion.Source: B. Hardware Validation · p. 5, Figure 3 · p. 6

V. CONCLUSION

• The framework integrates learned attachment-site discovery with existing assignment and collision-aware control for progressive assembly without a predefined deployment-time site list.The ablation reports improved morphology accuracy and consistency from fine-tuning over standard Growing NCA.Source: §V · p. 6