Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Safety-Aware Multi-Agent Learning for Dynamic Network Bridging

Raffaele Galliera, Konstantinos Mitsopoulos, Niranjan Suri, Raffaele Romagnoli

TL;DR

The paper addresses safe coordination of multiple agents to dynamically bridge communication between moving targets under partial observability. It proposes a decentralized safety filter based on ellipsoidal positively invariant sets $\mathcal{E}_c$ and augments learning with safety informed edge features and graph based message passing to improve coordination. Key contributions include a safe setpoint update algorithm with local one hop communication, explicit edge level safety signals, and empirical evidence that these features improve task performance while reducing safety interventions. The findings demonstrate that local safety enforcement coupled with decentralized learning can achieve effective and scalable coordination in safety critical distributed tasks.

Abstract

Addressing complex cooperative tasks in safety-critical environments poses significant challenges for multi-agent systems, especially under conditions of partial observability. We focus on a dynamic network bridging task, where agents must learn to maintain a communication path between two moving targets. To ensure safety during training and deployment, we integrate a control-theoretic safety filter that enforces collision avoidance through local setpoint updates. We develop and evaluate multi-agent reinforcement learning safety-informed message passing, showing that encoding safety filter activations as edge-level features improves coordination. The results suggest that local safety enforcement and decentralized learning can be effectively combined in distributed multi-agent tasks.

Safety-Aware Multi-Agent Learning for Dynamic Network Bridging

TL;DR

The paper addresses safe coordination of multiple agents to dynamically bridge communication between moving targets under partial observability. It proposes a decentralized safety filter based on ellipsoidal positively invariant sets and augments learning with safety informed edge features and graph based message passing to improve coordination. Key contributions include a safe setpoint update algorithm with local one hop communication, explicit edge level safety signals, and empirical evidence that these features improve task performance while reducing safety interventions. The findings demonstrate that local safety enforcement coupled with decentralized learning can achieve effective and scalable coordination in safety critical distributed tasks.

Abstract

Addressing complex cooperative tasks in safety-critical environments poses significant challenges for multi-agent systems, especially under conditions of partial observability. We focus on a dynamic network bridging task, where agents must learn to maintain a communication path between two moving targets. To ensure safety during training and deployment, we integrate a control-theoretic safety filter that enforces collision avoidance through local setpoint updates. We develop and evaluate multi-agent reinforcement learning safety-informed message passing, showing that encoding safety filter activations as edge-level features improves coordination. The results suggest that local safety enforcement and decentralized learning can be effectively combined in distributed multi-agent tasks.
Paper Structure (19 sections, 3 theorems, 21 equations, 5 figures, 1 table, 1 algorithm)

This paper contains 19 sections, 3 theorems, 21 equations, 5 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Task description
  4. Multi-agent reinforcement learning
  5. Distributed autonomous swarm formation for dynamic network bridging
  6. Agent Dynamics
  7. Safety
  8. Safety Filter for Tracking Control
  9. Decentralized Safety-Aware Agents
  10. Guiding learning via decentralized safety filters
  11. Safety-informed edges
  12. Policy parameterization
  13. Experimental Setup
  14. Dynamic network bridging environment implementation
  15. Training configurations
  16. ...and 4 more sections

Key Result

Proposition 1

Considering agent $i$ moving from target point $w_i$ to $w_{i+1}$, and updating the setpoint $x_{sp,i}$ every time $x_{i}(t)\in \mathcal{E}_s(x_{sp,i})$ by using update. Considering also Assumption 1 true, then the safety condition safe1 can be violated only with all the agents $j\neq i$ that satisf

Figures (5)

  • Figure 1: Decentralized swarm coordination for dynamic network bridging
  • Figure 2: Safe tracking control and safety verification.
  • Figure 3: A typical configuration of our system. Agents are represented with letters A, B, C and targets with T. Agent A and B have established a successful connection. Agent A receives node features $f_n^A$, $f_n^A$ and edge features $f_n^{AB}$ as inputs.
  • Figure 4: Impact of safety-informed edge features.
  • Figure 5: Sequence of actions executed by agents trained with Approach A when deployed in a LVC environment.

Theorems & Definitions (9)

  • Definition 1
  • Definition 2
  • Proposition 1
  • proof
  • Proposition 2
  • proof
  • Proposition 3
  • proof
  • Remark 1