Benchmark


Swarm Intelligence & Navigation Methodology Comparison

This benchmark evaluates three distinct paradigm approaches to swarm navigation and collision avoidance: traditional Finite State Machines (FSM), Reinforcement Learning (RL), and our proposed Spatioz architecture. Click the demo buttons above to visualize their behavior in real time.

Key Benchmark Metric Finite State Machine (FSM / Grid) Reinforcement Learning (RL / K-NN Continuous) Spatioz (Agentic Tension-Field Architecture)
Maximum Scalability Very Low (< 500 agents due to queue deadlocks). Low to Medium (< 2,000 agents due to model/tensor weight overhead). Very High (1 Million+ Agents).
Computational Complexity O(N) to O(N²) depending on grid search and collision detection resolution. O(N log N) for K-Nearest Neighbors + Feedforward Neural Network inference overhead. Highly Efficient based on optimized physics-inspired vector equations.
Rendering Performance (FPS) High (~60 FPS) due to simple deterministic conditional branches. Unstable / Drops significantly under high agent counts due to continuous model evaluation. High & Stable (60 FPS) optimized via lightweight latent space mapping.
Jittering & Movement Stutter Low jitter, but movements are rigid, sudden, and robotic. High micro-jittering and stuttering as agents continuously fluctuate between conflicting learned policies. Ultra-smooth continuous trajectory generation without erratic oscillations.
Emergent Patterns No emergent patterns. Agents act strictly in isolation, leading to chaotic individual behaviors. Disorganized "bubble-like" clustering patterns with agents constantly bumping and milling about. Dynamic Lane Formation. Agents naturally self-organize into structured flow lanes.
Swarm Density Tolerance Extremely low. Deadlocks occur frequently due to long, rigid sequential queues. Low. Swarm collapses into deadlock because agents lack the micro-space required to jitter or compute paths. High Density Resilience. Resolved using structural latent space state representation.
Behavioral Characteristics Rigid, deterministic, and highly artificial behavior. Tactical and opportunistic (navigating micro-gaps), yet highly erratic. Organic, flowing, and emergent collective behavior.
Hardware Requirements Very light CPU computational overhead. Very heavy (requires GPU/VRAM acceleration for medium scale). Light / Optimal (pure mathematical formulation).
Deadlock Resolution Requires manual recovery rules or arbitrary timeouts. Prone to systemic freezing under tight constraints. Self-Correcting. Structural latent space guides coordinated escape maneuvers.