Glossary

Terms used across Arboria Labs papers, tools, and foundation pages. Entries are short on purpose — each one links out to a longer treatment when there is one.

Swarm metrics

Cohesion. Mean distance of agents to their common centroid. Lower values mean a more tightly grouped swarm. Implemented in gossamer.utils.metrics.cohesion.

Alignment (ψ). Magnitude of the swarm’s mean unit-velocity vector, in [0, 1]. 1 means every agent is moving in the same direction, 0 means headings are uniformly distributed. Implemented in gossamer.utils.metrics.alignment.

Separation. Mean nearest-neighbor distance. Used as a proxy for crowding. Implemented in gossamer.utils.metrics.separation.

Coverage ratio (χ). Fraction of a gridded region that has been visited by any agent in the swarm at least once. Used as the canonical metric in the emergent-behavior paper.

Energy per kilogram (ε). For ISRU / construction scenarios, total energy consumed divided by mass produced. Used in the macro-micro synergy paper.

Collision rate (ρ_c). Collisions per agent per simulation step.

Distributed systems abstractions

CRDT. Conflict-free Replicated Data Type. State container whose merge operation is associative, commutative, and idempotent — so agents can reconcile divergent replicas without coordination. Core building block of the ICCD paper.

DTN. Delay-Tolerant Networking. Networking model for sparse connectivity, intermittent contacts, and hour-scale propagation delays. Bundles are stored and forwarded opportunistically. See RFC 4838 and successors .

AoI. Age of Information. Time elapsed since the last successful update of a piece of remote state. Optimizing AoI is different from optimizing throughput or latency; some bundles should be re-sent even when nothing new happened.

Contact plan. A scheduled list of upcoming (sender, receiver, start, duration) windows derived from predicted positions / orbits. Lets the DTN layer decide when to forward, not just whether to try.

Coordination and control

Consensus. Distributed agreement on a scalar or vector value via local exchanges. Laplacian average consensus is the baseline (consensus.py ); variants include Metropolis weights and event-triggered consensus.

Task allocation. Assignment of N tasks to M agents subject to a cost metric (usually distance or energy). Hungarian / auction / market algorithms all appear in the stack; see task_allocation.py .

Stigmergy. Indirect coordination via modifications to a shared environment — the mechanism behind ant colony optimization. See the ACO technique page and the TF-ACO paper.

ICCD. Intent-CRDT with Contact-Plan DTN. Arboria’s framework for maintaining mission-intent coherence at million-agent scale under hour-scale light delays.

HMA. Hierarchical Market Auction. Arboria’s two-tier auction mechanism for macro / micro coordination in planetary construction.

Modern multi-agent learning

MARL. Multi-Agent Reinforcement Learning.

CTDE. Centralized Training, Decentralized Execution. Parameters are trained with access to the full swarm state but deployed with local-only observations. Standard setup for scalable MARL.

MAPPO. Multi-Agent Proximal Policy Optimization. A shared-policy CTDE variant of PPO — a reference baseline across the Arboria Swarm Benchmark.

GNN. Graph Neural Network. Policy architecture that passes messages over the interaction graph; a natural fit for swarm problems because it handles variable agent counts and permutation invariance by construction.

JEPA. Joint-Embedding Predictive Architecture. World-model approach proposed by LeCun — learns to predict representations of future observations rather than the observations themselves. See modern techniques.

Dreamer. Recurrent world-model + policy from Hafner et al. The de facto baseline for model-based RL research.

Criticality and active matter

Vicsek model. Minimal continuous-time model of flocking (Vicsek et al., 1995). Identifies a noise-driven phase transition between disordered motion and coherent flocking — the universality class most flocking claims are measured against.

Order parameter. Macroscopic observable whose value distinguishes ordered from disordered phases. Alignment ψ is the standard choice for flocking.

Susceptibility (χ). Variance of the order parameter (scaled by system size). Peaks at the critical point; one of the primary tools for locating phase transitions empirically.

Binder cumulant (U). 1 - <ψ⁴>/(3<ψ²>²). Finite-size-scaling tool — curves at different system sizes cross at the critical point.

Branching ratio (σ). <n_{t+1}> / <n_t> for an event train. σ = 1 is the signature of a critical process (self-organized criticality); useful for cascade / avalanche analysis.

Transfer entropy. Information-theoretic measure of directed influence from one time series to another, beyond what history predicts. See gossamer.metrics.info.