VORL-EXPLORE: A Hybrid Learning Planning Approach to Multi-Robot Exploration in Dynamic Environments

Ning Liu; Sen Shen; Zheng Li; Sheng Liu; Dongkun Han; Shangke Lyu; Thomas Braunl

cs.ROMar 9, 2026

VORL-EXPLORE: A Hybrid Learning Planning Approach to Multi-Robot Exploration in Dynamic Environments

Authors: Ning Liu, Sen Shen, Zheng Li, Sheng Liu, Dongkun Han, Shangke Lyu, Thomas Braunl

Organizations: The University of Western Australia · The Chinese University of Hong Kong · Karlsruhe Institute of Technology · Nanjing University

Abstract

Hierarchical multi-robot exploration commonly decouples frontier allocation from local navigation, which can make the system brittle in dense and dynamic environments. Because the allocator lacks direct awareness of execution difficulty, robots may cluster at bottlenecks, trigger oscillatory replanning, and generate redundant coverage. We propose VORL-EXPLORE, a hybrid learning and planning framework that addresses this limitation through execution fidelity, a shared estimate of local navigability that couples task allocation with motion execution. This fidelity signal is incorporated into a fidelity-coupled Voronoi objective with inter-robot repulsion to reduce contention before it emerges. It also drives a risk-aware adaptive arbitration mechanism between global A* guidance and a reactive reinforcement learning policy, balancing long-range efficiency with safe interaction in confined spaces. The framework further supports online self-supervised recalibration of the fidelity model using pseudo-labels derived from recent progress and safety outcomes, enabling adaptation to non-stationary obstacles without manual risk tuning. We evaluate this capability separately in a dedicated severe-traffic ablation. Extensive experiments in randomized grids and a Gazebo factory scenario show high success rates, shorter path length, lower overlap, and robust collision avoidance. The source code will be made publicly available upon acceptance.

Figures & tables

Fig. 1 : Canonical Voronoi-based frontier assignment in multi-robot exploration. Three robots share an occupancy map, extract frontiers, and form a Voronoi partition $\{V_{i,t}\}$ induced by BFS distance. Each robot selects a frontier within its region as the next exploration goal.

Fig. 2 : Closed-loop architecture of VORL-EXPLORE . Each robot estimates execution fidelity $p_{i,t}$ online from local cues and shared teammate states. In the task layer, $p_{i,t}$ modulates frontier-scoring weights to reduce assignments that are likely to cause congestion. In the motion layer, $p_{i,t}$ drives a hysteresis gate that selects between $A^{*}$ guidance and a reactive RL policy. Progress and safety outcomes generate pseudo-labels to update the fidelity estimator online, closing the coupling loop.

${\color[rgb]{0,0.58,0.42}\bullet}$ 40 $\times$ 40 world, 4 Agents, 30% static obstacle density
Method	8 Dynamic obstacles			16 Dynamic obstacles			32 Dynamic obstacles			64 Dynamic obstacles
Method	SR $\uparrow$	EL $\downarrow$	Overlap $\downarrow$	SR $\uparrow$	EL $\downarrow$	Overlap $\downarrow$	SR $\uparrow$	EL $\downarrow$	Overlap $\downarrow$	SR $\uparrow$	EL $\downarrow$	Overlap $\downarrow$
DHC	0.90 $\pm$ 0.08	169.22 $\pm$ 11.3	0.23 $\pm$ 0.03	0.88 $\pm$ 0.09	171.78 $\pm$ 12.1	0.23 $\pm$ 0.03	0.87 $\pm$ 0.10	179.64 $\pm$ 14.5	0.23 $\pm$ 0.03	0.88 $\pm$ 0.09	191.45 $\pm$ 15.2	0.24 $\pm$ 0.04
PICO	0.71 $\pm$ 0.12	212.80 $\pm$ 18.4	0.32 $\pm$ 0.05	0.62 $\pm$ 0.15	213.60 $\pm$ 19.1	0.32 $\pm$ 0.05	0.56 $\pm$ 0.14	221.40 $\pm$ 20.3	0.33 $\pm$ 0.06	0.51 $\pm$ 0.16	231.42 $\pm$ 21.5	0.33 $\pm$ 0.06
MATS-LP	0.85 $\pm$ 0.09	198.75 $\pm$ 14.2	0.26 $\pm$ 0.04	0.82 $\pm$ 0.10	203.12 $\pm$ 15.4	0.26 $\pm$ 0.04	0.85 $\pm$ 0.11	205.33 $\pm$ 16.1	0.27 $\pm$ 0.05	0.87 $\pm$ 0.10	207.84 $\pm$ 17.5	0.25 $\pm$ 0.04
ICBS	0.78 $\pm$ 0.11	182.20 $\pm$ 13.1	0.27 $\pm$ 0.04	0.68 $\pm$ 0.14	194.40 $\pm$ 16.5	0.29 $\pm$ 0.05	0.48 $\pm$ 0.16	218.30 $\pm$ 22.4	0.35 $\pm$ 0.07	0.22 $\pm$ 0.18	235.60 $\pm$ 30.2	0.41 $\pm$ 0.09

TABLE I : Experimental results on

40\times 40

and

80\times 80

grids over 100 runs. Values are reported as mean

\pm

standard deviation. SR: success rate (

\uparrow

), EL: Exploration length (

\downarrow

), and Overlap: redundant overlap ratio (

\downarrow

). Best values are bolded .

Panel of Figure 4 — Fig. 3 : Allocator-level comparison with target allocation baselines on an $80\times 80$ grid with 16 dynamic obstacles and 30% static obstacle density. All methods share the same VORL-Explore execution layer. Curves show the mean over 100 runs and shaded regions indicate 95% confidence intervals.

Panel of Figure 5 — Fig. 4 : Progressive ablation of the coupled architecture. Base disables both coupling links and uses decoupled target assignment and motion execution. CA enables fidelity-coupled assignment by using execution fidelity to reweight frontier scoring, while keeping the execution module unchanged. CP enables fidelity-gated switching in the execution layer, while keeping the assignment objective unchanged. Full enables both CA and CP. Bars show the mean and error bars indicate $\pm$ STD over 100 runs with 4 agents on a $40\times 40$ grid, 30% static obstacle density, and 32 dynamic obstacles.

Variant	Setup		Metrics
Variant	Pre-train	Online	SR $\uparrow$	EL $\downarrow$	Recov. $\downarrow$	$\mathbf{\pi_{plan}}$ (%)
$\text{VORL}_{\text{Static}}$ (Cold)	$\times$	$\times$	0.36 $\pm$ 0.11	248.2 $\pm$ 29.5	82.4 $\pm$ 15.2	91.5
$\text{VORL}_{\text{Static}}$ (Warm)	✓	$\times$	0.51 $\pm$ 0.13	224.5 $\pm$ 22.1	42.7 $\pm$ 9.6	74.2
$\text{VORL}_{\text{Adapt}}$ (Cold)	$\times$	✓	0.65 $\pm$ 0.07	208.7 $\pm$ 17.5	12.3 $\pm$ 4.2	37.8
VORL-EXPLORE	✓	✓	0.69 $\pm$ 0.05	204.3 $\pm$ 14.2	6.8 $\pm$ 2.1	34.5

TABLE II : Ablation of gate initialization (Pre-train) and test-time gate updates (Online) under severe traffic. Values are mean

\pm

STD over 100 runs on a

40\times 40

grid with 30% static obstacles and 128 dynamic obstacles.

Panel of Figure 7 — Fig. 5 : Gazebo validation with four Pioneer3 robots in a dynamic factory environment. The normalized new coverage curve shows faster exploration than ROS explore_lite.

Explore similar work

SEAMLiS: Visibility-Aware Safety for Perception-Limited Multi-Robot Exploration
Jul 10, 2026Taekyung Kim, Rahul H Kumar, Aswin D. Menon +2Robotic Control Multi-Robot Exploration
Department of Robotics, University of Michigan, Ann Arbor, MI, 48109, USA · Department of Aerospace Engineering, University of Michigan, Ann Arbor, MI, 48109, USA · Department of Mechanical Engineering, University of Michigan, Ann Arbor, MI, 48109, USA
Cooperative Risk-Aware Exploration in Heterogeneous Multi-Robot Systems Using Algorithmic Altruism
Aug 28, 2026Brooks A. Butler, Jair Certório, João P. Hespanha +1Multi-Robot Exploration Multi-Robot Motion Planning
School of Electrical and Computer Engineering at Oklahoma State University, Stillwater, OK, 74074, USA · Systems Engineering Division, Instituto Tecnológico de Aeronáutica, Fortaleza, CE, 60415-513, Brazil · Electrical and Computer Engineering Department, University of California, Santa Barbara, CA, 93106, USA +1
Navigating the Clutter: Waypoint-Based Bi-Level Planning for Multi-Robot Systems
Apr 22, 2026Jiabao Ji, Yongchao Chen, Yang Zhang +4RL for Robotics Multi-Robot Motion Planning
UC Santa Barbara · Massachusetts Institute of Technology · Harvard University +2

Jul 10, 2026J/K move · Enter open · S save

SEAMLiS: Visibility-Aware Safety for Perception-Limited Multi-Robot Exploration

Taekyung Kim, Rahul H Kumar, Aswin D. Menon, Tzu-Hsiang Lin+1

Department of Robotics, University of Michigan, Ann Arbor, MI, 48109, USA · Department of Aerospace Engineering, University of Michigan, Ann Arbor, MI, 48109, USA · Department of Mechanical Engineering, University of Michigan, Ann Arbor, MI, 48109, USA

Robotic Control Multi-Robot Exploration Safe Robot Navigation Robot Safety Robot Collision Avoidance

Open paper AskPDF arXiv

Autonomous exploration in unknown environments is typically driven by informative frontiers, viewpoints, or trajectories, while local safety controllers avoid obstacles represented in the current map. Under finite sensing range and limited field of view, this separation can be unsafe: an exploration stack may plan optimistically through unobserved space and steer the sensor toward information gain rather than along the direction of motion, causing hidden obstacles to be detected too late for bounded-actuation avoidance. This paper presents SEAMLiS (Safe Exploration for Autonomous Multi-Robot Systems Under Limited Sensing), a modular execution-layer safety framework for decentralized multi-robot exploration. SEAMLiS preserves the upstream exploration stack, including the goal allocator and local planner, and enforces safety at the execution layer through perception-aware attitude and positional filters. A gatekeeper-based attitude filter switches between a visibility-promoting yaw policy and a velocity-tracking backup policy to preserve visibility of the critical known-free/unknown boundary with sufficient braking margin. A Control Barrier Function (CBF)-based positional filter then avoids known obstacles, newly detected obstacles, and other robots. We provide sufficient collision-avoidance conditions and validate the framework in randomized simulation, Isaac Sim, and Crazyflie hardware experiments. Results show collision-free exploration across tested single- and multi-robot settings while retaining much of the efficiency of visibility-promoting yaw control.

Abstract

Explore similar work

SEAMLiS: Visibility-Aware Safety for Perception-Limited Multi-Robot Exploration

Cooperative Risk-Aware Exploration in Heterogeneous Multi-Robot Systems Using Algorithmic Altruism

Navigating the Clutter: Waypoint-Based Bi-Level Planning for Multi-Robot Systems