Resilient Decentralized Shape Formation of Multi-Robot Swarms Under Adversarial Real-World Faults in High-Complexity Disaster Environments

Decentralized multi-robot shape formation in GPS-denied environments is important for search-and-rescue and disaster response. Conventional consensus-based formation controllers can violate safe separation, oscillate, or stall when communication degrades, sensors are noisy, or an agent drops out. This paper proposes a resilient decentralized controller that combines Huber-weighted consensus and Exponential Moving Average (EMA) state estimation with a Dual-Layer Artificial Potential Field (APF) safety barrier. The first layer adds an anti-symmetric lateral deflection to resolve head-on encounters between peers; the second applies a normal-plus-tangential circulation field around static obstacles. The framework is evaluated on five simulated nonholonomic differential-drive robots in Gazebo Harmonic (ROS 2 Jazzy). Three runs are compared: a linear consensus baseline, the same baseline under faults (15% packet loss, 50-150 ms delay, σ = 0.18 m sensor noise, and a mid-run agent blackout), and the proposed controller under the same faults. The fault-free baseline reached the pentagon fastest (0.0156 m final RMS error) but its minimum separation fell to 0.3918 m, with 31 logged violations of the 0.50 m margin. Under faults the baseline reached 0.0546 m with 5 violations (minimum 0.4833 m). The proposed controller recorded no violations, kept a minimum separation of 0.6757 m, and reached 0.0368 m final error, at the cost of slower convergence (about 21 s versus about 12-15 s). Results are from single simulation runs, and the safety barrier is empirical rather than formally guaranteed.