We present a communication-free method for safe multi-robot coordination in complex environments such as forests with dense canopy cover, where GNSS is unavailable. Our approach relies on an onboard anisotropic 3D LiDAR sensor used for SLAM as well as for detecting obstacles and neighboring robots. We develop a novel perception-aware 3D navigation framework that enables robots to safely and effectively progress toward a goal region despite limited sensor field-of-view. The approach is evaluated through extensive simulations across diverse scenarios and validated in real-world field experiments, demonstrating its scalability, robustness, and reliability.
Figures & tables
Figure 1: Snapshot during a real-world experiment in the forest with 3 UAVs.
EGO2 [ 3 ]
RMADER [ 5 ]
PACNAV [ 6 ]
RBL [ 4 ]
HDSM [ 7 ]
RVC-NMPC [ 8 ]
Chen et al. [ 9 ]
DREAM [ 10 ]
AMSWARM [ 11 ]
Choi et al. [ 12 ]
iRBL (our)
Table I : Qualitative comparison with state-of-the-art algorithms on the aspects of interest (i.e., implicit coordination ( ), anisotropic perception ( ), test in the wild ( ) and operational speed ( )). We indicate with “ ” if the algorithm does not implement the feature, “ ” if it does, and “ ” if it does only partially. In the last row the operational speed is indicated with arrows, where “ ” indicates poor performance, “ ” indicates high performance, while “ ” and “ ” indicate intermediate low and intermediate high performance respectively.
Figure 2: Graphical representation of the sensor field-of-view FoV={fx,fz,fa}\penalty(deg) .
Figure 3: 2D Graphical representation of the set Ai . On the left, the case where the robots are separated by more than 2Δij ; on the right, the case where they are separated by less than 2Δij . The half-spaces are constructed according to ( 1 ) such that all the points are collision-free.
Figure 4: 2D Graphical representation of the set Ci (which in this particular case correspond also with the set Bi ) computed using the proposed seeds. The obstacles are represented by grey circles.
Figure 5: Illustration of the main variables associated with the i -th robot during a real-world experiment. The robot’s pose pi is shown as an axis reference frame. The set Wi is represented by the magenta set, and the centroid position cBi is shown in orange. The planned path is depicted by a red line. The wpi and ei indicate respectively the active waypooint and the final goal positions. Other robots positions pj are indicated by a grey box. The colored point cloud illustrates obstacles, with different colors corresponding to different z-values.
Figure 6: Block diagram representation of the proposed pipeline.
Figure 7: Top view of the robot’s trajectory using iRBL in an environment with low traversability ( τ=1.57 ), where EGO2 [ 17 ] and RBL [ 4 ] are unable to successfully navigate. The color scale represents the robot’s velocity, as indicated by the colorbar.
τ (m)
l (m)
t (s)
vˉ (m/s)
dmino (m)
SR
EGO2 [ 17 ]
3.566
31.37
15.17
2.07
0.486
0.4 ∣ 1 ∣ 1
2.036
36.01
20.32
1.82
0.46
0 ∣ 1 ∣ 1
1.142
35.25
23.20
1.51
0.52
0.2 ∣ 0.2 ∣ 0.8
0.984
-
-
-
-
0.0 ∣ 0.0 ∣ 0.8
RBL [ 4 ]
3.566
36.31
52.9
0.68
0.45
1 ∣ 0.6 ∣ 1
2.036
34.56
71.92
0.48
0.43
1 ∣ 0.4 ∣ 1
Table II : Comparison with state-of-the-art. Simulation with N=1 robot and δi=0.2 m in 4 different cluttered environments. We report the traversability τ , the length of the path l , the time to reach the goal t , the average speed vˉ , the minimum distance to obstacles dmino , and the success rate SR in terms of violation of acceleration limits, covergence to the goal, and safety ( dmino≥δi ). We repeat each scenario 10 times, and report the average results. All the parameters are kept the same for all the different scenarios.
Figure 8: Top view for the robots’ trajectories with the iRBL half FoV case and δi=0.2 (m), in three different scenarios. The crossing circle without obstacle ( N=10 ), crossing circle with obstacles ( N=10 ) and forest-like environment ( N=5 ). The color indicates the velocities according to the colorbars.
τ (m)
δi (m)
l (m)
t (s)
vˉ (m/s)
dmin (m)
iRBL (lim.FoV)
-
0.2
33.14
23.43
1.41
0.737
-
0.5
33.99
26.72
1.26
1.155
-
1.0
35.51
33.95
1.04
2.17
iRBL (half FoV)
-
0.2
33.56
22.14
1.50
0.729
-
0.5
34.30
25.09
1.35
1.402
-
1.0
35.14
27.62
1.26
2.092
Table III : Simulation results with N=10 robots in a crossing circle scenario without obstacles. We report the length of the path l , the time to reach the goal t , the average speed vˉ , and the minimum distance between the robots dmin . We considered 4 different field-of-views (FoV). The limited FoV {180,59,-20} (deg), half FoV {180,180,-90} (deg), full FoV {180,360,0} (deg), and the two-dimensional case {360,0,0} (deg).
τ (m)
δi (m)
l (m)
t (s)
vˉ (m/s)
dmin (m)
iRBL (lim. FoV)
6.56
0.2
32.46
23.24
1.40
0.661
4.60
0.5
35.83
33.07
1.08
1.165
2.21
1.0
38.54
55.06
0.69
2.199
iRBL (half FoV)
6.56
0.2
32.70
23.74
1.37
0.878
4.60
0.5
35.99
34.47
1.04
1.178
2.21
1.0
38.17
43.75
0.86
2.119
Table IV : Simulation results with N=10 robots in a crossing circle scenario with obstacles. We report the traversability τ , the length of the path l , the time to reach the goal t , the average speed vˉ , and the minimum distance between the robots dmin . We considered 4 different field-of-views (FoV). The limited FoV {180,59,-20} (deg), half FoV {180,180,-90} (deg), full FoV {180,360,0} (deg), and the two-dimensional case {360,0,0} (deg).
τ (m)
δi (m)
l (m)
t (s)
vˉ (m/s)
dmin (m)
iRBL (lim. FoV)
8.67
0.2
34.58
26.83
1.28
0.976
7.01
0.5
36.51
30.94
1.18
1.723
4.17
1.0
39.60
52.31
0.76
2.216
iRBL (half FoV)
8.67
0.2
35.10
26.52
1.33
1.151
7.01
0.5
36.94
30.25
1.22
1.768
4.17
1.0
37.31
34.77
1.07
2.235
Table V : Simulation results with N=5 robots in a forest-like environment. We report the traversability τ , the length of the path l , the time to reach the goal t , the average speed vˉ , and the minimum distance between the robots dmin . We considered 4 different field-of-views (FoV). The limited FoV {180,59,-20} (deg), half FoV {180,180,-90} (deg), full FoV {180,360,0} (deg), and the two-dimensional case {360,0,0} (deg).
Figure 9: Different experimental scenarios. The top row contains snapshots from the camera, while the bottom row shows the corresponding snapshots extracted from RViz for the same experiment. Similarly to Figure 5 , the robot pose is represented by an axis reference frame. The set Wi is depicted as blue dots, and the centroid cBi as an orange sphere. The planned path is shown as a red line. The active waypoint wpi lies on the planned path and is represented by a purple sphere. The grey boxes represent the positions of the other robots. The colored point cloud represents obstacles, where different colors correspond to different z -values. The “eye” symbol indicates the robot whose viewpoint is currently shown in the bottom RViz visualization.
Figure 10: Different viewpoints from experiment #1 in Table VI . We depict the trajectories followed by the two robots during the mission. The robots’ velocities can be inferred from the color-bars on the right-hand side. Different colors in the point cloud indicate different z -values.
scenario
uav
τ (m)
l (m)
t (s)
vˉ (m/s)
vmax (m/s)
#1 (40m)
1
3.10
45.77
23.00
1.99
2.60
2
3.10
41.98
23.45
1.79
2.85
#2 (80m)
1
16.02
86.97
35.07
2.48
2.79
2
16.02
91.02
45.51
2.00
2.69
3
16.02
94.34
53.00
1.78
2.81
#3 (40m)
1
6.73
46.06
23.03
2.00
3.32
Table VI : Quantitative results of experiments in three different scenarios. All the scenarios involve UAVs starting from close positions and with close goal locations. For all the scenarios we report the traversability τ , the length of the path l , the time to reach the goal t , the average speed vˉ , and the maximum velocity achieved vmax . The mission environments of scenario 1,2 and 3 are illustrated respectively in Figure 9 -(a),(b) and (c).
The LiDAR-based multi-agent and single-agent perception has shown promising performance in environmental understanding for robots and automated vehicles. However, there is no existing method that simultaneously solves both multi-agent and single-agent perception in an unsupervised way. By sharing sensor data between multiple agents via communication, this paper discovers two key insights: 1) Improved point cloud density after the data sharing from cooperative views could benefit unsupervised object classification, 2) Cooperative view of multiple agents can be used as unsupervised guidance for the 3D object detection in the single view. Based on these two discovered insights, we propose an Unsupervised Multi-agent and Single-agent (UMS) perception framework that leverages multi-agent cooperation without human annotations to simultaneously solve multi-agent and single-agent perception. UMS combines a learning-based Proposal Purifying Filter to better classify the candidate proposals after multi-agent point cloud density cooperation, followed by a Progressive Proposal Stabilizing module to yield reliable pseudo labels by the easy-to-hard curriculum learning. Furthermore, we design a Cross-View Consensus Learning to use multi-agent cooperative view to guide detection in single-agent view. Experimental results on two public datasets V2V4Real and OPV2V show that our UMS method achieved significantly higher 3D detection performance than the state-of-the-art methods on both multi-agent and single-agent perception tasks in an unsupervised setting.
Collaborative navigation of heterogeneous robots in unknown environments poses significant challenges due to sensing, communication, and computational limitations. In this work, a lead robot navigates toward a target while a mobile sensor robot (e.g., a drone) assists by transmitting information about its locally observed map under bandwidth constraints. We propose a framework that enables the sensor to jointly select its transmitted map points and navigation actions online, while also predicting unexplored regions of the environment. To this end, we present β-Sparse Gaussian Processes, a robust variational sparse Gaussian Process model for task-aware inducing point selection under cardinality constraints. Furthermore, we develop an action-selection strategy that balances task relevance with exploration. Simulations on Mars and Earth maps show that the framework can reduce path cost by 18% relative to no communication and decrease transmitted information by 76% compared to raw-data transmission baselines.
D. Guggenheim School of Aerospace Engineering, Georgia Tech, Atlanta, GA, USA · Department of Electrical and Computer Engineering, UNC Charlotte, Charlotte, NC, USA
Multi-robot systems rely on underlying connectivity to ensure reliable communication and timely coordination. This paper studies the line-of-sight (LoS) connectivity maintenance problem in multi-robot navigation with unknown obstacles. Prior works typically assume known environment maps to formulate LoS constraints between robots, which hinders their practical deployment. To overcome this limitation, we propose an inherently distributed approach where each robot only constructs an egocentric visible region based on its real-time LiDAR scans, instead of endeavoring to build a global map online. The individual visible regions are shared through distributed communication to establish inter-robot LoS constraints, which are then incorporated into a multi-robot navigation framework to ensure LoS-connectivity. Moreover, we enhance the robustness of connectivity maintenance by proposing a more accurate LoS-distance metric, which further enables flexible topology optimization that eliminates redundant and effort-demanding connections. The proposed framework is evaluated through extensive multi-robot navigation and exploration tasks in both simulation and real-world experiments. Results show that it reliably maintains LoS-connectivity between robots in challenging environments cluttered with obstacles, even under large visible ranges and fragile minimal topologies, where existing methods consistently fail. Ablation studies also reveal that topology optimization boosts navigation efficiency by around 20%, demonstrating the framework's potential for efficient navigation under connectivity constraints.