Topology-Driven Anti-Entanglement Control for Soft Robots

Soft robots are increasingly vital for precision manufacturing in constrained environments, but coordinating multiple robots to avoid entanglement remains a core challenge. Existing distributed training frameworks struggle with observability in high-density barriers and unstable environments, leading to poor learning. To solve this, researchers propose a Topology-Driven Multi-Agent Reinforcement Learning (TD-MARL) framework that uses topological invariants to assess and mitigate entanglement risk. The architecture includes a centralized critic that shares topological states across agents, stabilizing training without requiring inter-robot communication during deployment. A dedicated topological security layer prevents policies from falling into local optima by continuously evaluating entanglement risk using geometric invariants.

Full simulation experiments in realistic environments show TD-MARL achieves faster convergence and better anti-winding performance compared to state-of-the-art deep reinforcement learning (DRL) methods. The framework's distributed execution eliminates communication overhead, improving system reliability in bandwidth-constrained industrial settings. By formalizing entanglement risk through topology rather than heuristic rules, the approach offers a scalable path for deploying multi-robot systems in complex manufacturing, search-and-rescue, and medical applications where physical tangling can cause catastrophic failures. The paper includes 17 pages and 4 figures, available on arXiv (2605.05236).