GIANT - Global Path Integration and Attentive Graph Networks for Multi-Agent Trajectory Planning
Authors: Jonas le Fevre Sejersen, Toyotaro Suzumura, Erdal Kayacan
Organizations: Artificial Intelligence in Robotics Laboratory (AiR Lab), Department of Electrical and Computer Engineering, Aarhus University, 8000 Aarhus C, Denmark · Foundation Models for Artificial Intelligence group, Department of Information and Communication Engineering, Tokyo University, Tokyo, Japan · Automatic Control Group, Department of Electrical Engineering and Information Technology, Paderborn University, Paderborn, Germany
This paper presents a novel approach to multi-robot collision avoidance that integrates global path planning with local navigation strategies, utilizing attentive graph neural networks to manage dynamic interactions among agents. We introduce a local navigation model that leverages pre-planned global paths, allowing robots to adhere to optimal routes while dynamically adjusting to environmental changes. The models robustness is enhanced through the introduction of noise during training, resulting in superior performance in complex, dynamic environments. Our approach is evaluated against established baselines, including NH-ORCA, DRL-NAV, and GA3C-CADRL, across various structurally diverse simulated scenarios. The results demonstrate that our model achieves consistently higher success rates, lower collision rates, and more efficient navigation, particularly in challenging scenarios where baseline models struggle. This work offers an advancement in multi-robot navigation, with implications for robust performance in complex, dynamic environments with varying degrees of complexity, such as those encountered in logistics, where adaptability is essential for accommodating unforeseen obstacles and unpredictable changes.
Figures & tables
Figure 1: The figure illustrates the trajectory chosen by our model, where the robot navigates through highly dense and dynamic environments while adhering to the global path
Figure 2: Overall illustration of the observation space. (a) Illustrates the environmental obstacles captured by the LiDAR, the velocity of the agent, and the goal distance. (b) Illustrates the global path observations, with the distance from the agent (red) to the target point (blue), the angle difference between the agent heading and the target point, and the direction of the global path at the target point. (c) Illustrates the estimated dynamic clusters extracted from the LiDAR. The position and velocity of each cluster are tracked over time and fed to the network as a graph.
Figure 3: An illustration of the actor-network architecture. The network includes three observation encoders: a static LiDAR encoder that processes the current frame through Conv1D layers to capture static features, a temporal LiDAR encoder that processes the last three frames through Conv1D layers to learn dynamic features, and a neighbor attentive Graph Neural Network encoder. The combined encoded features are passed to fully connected layers and the network outputs the mean and standard deviation for the action normal distributions, which include linear and angular velocities. The critic network follows an identical structure but outputs state values instead.
Parameters
Values
Parameters
Values
learning rate critic
4e-4
Δt
0.1s
learning rate actor
2e-5
H
5
entropy coefficient
0
rp
2.5
reward discount
0.99
rs
-0.25m
GAE discount
0.95
PS
0.3m
PPO train epoch
10
rgoal
15
Table I: Training hyperparameters
Figure 4: The six different environments used during training.
Metric
Full Model
No Global Path
No GNN
Success Rate (%)
96.7
92.7
93.2
Stuck / Collision Rate (%)
3.3 / 0.
7.1 / 0.2
2.2 / 4.6
Extra Time (Normalized)
18.7
21.8
14.7
Average Speed (m/s)
0.42
0.34
0.57
Table II: Ablation Study Results in the Doorway Environment with 15 Agents.
Training settings
Test environments with variable number of agents:
Metrics
Methods
Circle (radius 7.5m)
Doorway
Narrow Corridor
Random (8 obstacles)
10
20
40
5
10
15
8
12
16
10
20
40
NH-ORCA [ 12 ]
72.5
32.7
91.67
93
95.5
97.25
DRL-Nav [ 1 ]
91.3
93.8
82.5
95.3
GA3C-CADRL [ 4 ]
52.5
14.7
73.75
53.17
83.8
Success Rate (%)
GIANT [ours]
96.67
97.5
99.5
99.25
Table III: Model performance on evaluation scenarios
This work enhances global path planning via a pure-pursuit controller with multi-model kinematic switching that sustains plan fidelity across diverse terrains. The system includes a traversability graph for terrain analysis, a Heading-Aware A* algorithm for generating feasible paths, and a multi-model Pure Pursuit controller for dynamic tracking. A core innovation is adaptive kinematic modeling, enabling real-time switching between kinematic models based on terrain features and robot states. This adaptability optimizes path efficiency and energy use in challenging scenarios. We validate the approach in simulation on different platforms, namely the Artaban quadruped and the X3 quadrotor drone, showcasing improved performance, robustness, and adaptability over standard baselines.
Pietro Gori, Francesco Iotti, Eduard Zelenay +5
Centro di Ricerca “Enrico Piaggio” and the Dipartimento di Ingegneria dell’Informazione, Universit`a di Pisa, Largo Lucio Lazzarino 1, 56126 Pisa, Italy · Institute of Robotics and Cybernetics, Faculty of Electrical Engineering and Information Technology, Slovak University of Technology in Bratislava, Ilkovicova 3, 812 19 Bratislava, Slovakia · Panza Robotics, National Center of Robotics, Ilkovicova 3, 841 04 Bratislava, Slovakia +1
The Collaborative Task Sequencing and Multi-Agent Path Finding (CTS-MAPF) problem requires agents to accomplish sequences of tasks while avoiding collisions, posing significant challenges due to its combinatorial complexity. This work introduces CTS-PLL, a hierarchical framework that extends the configuration-based CTS-MAPF planning paradigm with two key enhancements: a lock agents detection and release mechanism leveraging a complete planning method for local re-planning, and an anytime refinement procedure based on Large Neighborhood Search (LNS). These additions ensure robustness in dense environments and enable continuous improvement of solution quality. Extensive evaluations across sparse and dense benchmarks demonstrate that CTS-PLL achieves higher success rates and solution quality compared with existing methods, while maintaining competitive runtime efficiency. Real-world robot experiments further demonstrate the feasibility of the approach in practice.
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.
Ning Liu, Sen Shen, Zheng Li +4
The University of Western Australia · The Chinese University of Hong Kong · Karlsruhe Institute of Technology +1