DSGN: A Dual-Stream Gated Network for Scalable Simultaneous Surveillance and Containment

Heng Zhang; Ziyang Zuo; Guoxiang Zhao; Xiaoqiang Ren

cs.ROOct 25, 2026

DSGN: A Dual-Stream Gated Network for Scalable Simultaneous Surveillance and Containment

Authors: Heng Zhang, Ziyang Zuo, Guoxiang Zhao, Xiaoqiang Ren

Organizations: School of Mechatronic Engineering and Automation, Shanghai University, Shanghai 200444, China · School of Future Technology, Shanghai University, Shanghai 200444, China · National Key Laboratory of Maritime Unmanned Swarm Intelligence, Shanghai University, Shanghai 200444, China

Abstract

The simultaneous surveillance and containment (SSC) problem arises in multi-robot applications such as perimeter security and environmental monitoring, where robots must survey a region of interest and coordinate to enclose targets once they are detected. Surveillance favors spatial dispersion for coverage, whereas containment requires target-centric cohesion for enclosure. This conflict makes heuristic switching unstable, and the instability becomes more pronounced as larger robotic networks introduce more inter-agent and agent-target relations. End-to-end DRL reduces reliance on hand-crafted switching, but conventional single-stream policies often mix surveillance and containment cues in one shared feature space that could entangle dispersion-oriented and target-centric signals and weaken task transitions. In this paper, we propose the Dual-Stream Gated Network (DSGN), an architecture designed to address scale-amplified task-transition difficulties by separating surveillance and containment into specialized representation streams and coordinating them through a context-aware gating mechanism. A probabilistic prior further regularizes the gating process, reducing ambiguous intermediate decisions and supporting clearer behavioral transitions under large-scale interaction complexity. Evaluations show that DSGN maintains strong task efficacy and scalability under constant-density cross-scale settings. Additional ablation and phase-specific evaluations show that the proposed dynamic gating mechanism and probabilistic prior support effective task arbitration and reliable mode switching in both pure surveillance and task-switching scenarios. The source code is available at https://github.com/zuozysama/DSGN.

Figures & tables

Panel of Figure 1 — Fig. 1: Visualizing the dual kinematic objectives.

Fig. 2: The proposed DSGN architecture. Type-specific observation features are encoded and fused through dual-stream attention and a vector gate to produce actions.

Panel of Figure 3 — Fig. 3: State-dependent Beta priors for gate regularization.

Training Steps ( $10^{6}$ )	Pursuers	Evaders	Obstacles
$0\leq T_{\text{train}}<2$	3	1	0
$2\leq T_{\text{train}}<4$	4	1	1
$4\leq T_{\text{train}}<5$	7	2	2
$5\leq T_{\text{train}}<6$	11	3	4
$6\leq T_{\text{train}}\leq 7$	15	4	6

TABLE I: Progressive Curriculum Schedule

Fig. 4: Arbitration Variants across Nine Scale–Density Scenarios. (a) Strict SR; (b)

T_{\mathrm{cap}}

(s); (c) CR (log scale); (d) P-Steps. Error bars show 95% confidence intervals across 5 independent training seeds, with 40 trials per seed and scenario.

Map Scale	Region Size (m $\times$ m)	Pursuers	Evaders	Obstacles	Density Level $\boldsymbol{\rho_{m}}$ ( $10^{-3}$ m -2 )
Small	$112\times 112$	12	3	3	Low (1.196)
		16	4	3	Medium (1.594)
		20	5	3	High (1.993)
Medium	$200\times 200$	38	10	10	Low (1.200)
		51	13	10	Medium (1.600)
		64	16	10	High (2.000)

TABLE II: Configurations for Cross-Scale Evaluation

Method	3v1	6v2	9v3
Heuristic	0.000 / 488.4 [0.000, 0.000] / [465.0, 511.9]	0.000 / 500.0 [0.000, 0.000] / [500.0, 500.0]	0.000 / 500.0 [0.000, 0.000] / [500.0, 500.0]
DQN	0.000 / 500.0 [0.000, 0.000] / [500.0, 500.0]	0.000 / 500.0 [0.000, 0.000] / [500.0, 500.0]	0.000 / 500.0 [0.000, 0.000] / [500.0, 500.0]
IQN	0.000 / 451.3 [0.000, 0.000] / [415.4, 487.1]	0.950 / 169.5 [0.879, 1.000] / [155.4, 183.6]	0.785 / 112.5 [0.650, 0.920] / [106.7, 118.3]
TransIQN	0.025 / 455.0 [0.000, 0.076] / [413.8, 496.2]	0.840 / 190.8 [0.724, 0.956] / [160.6, 220.9]	0.725 / 103.9 [0.580, 0.870] / [84.8, 122.9]
Random	0.000 / 467.2 [0.000, 0.000] / [429.6, 504.8]	0.075 / 475.8 [0.000, 0.160] / [450.3, 501.4]	0.300 / 405.7 [0.152, 0.448] / [359.3, 452.1]
Average	0.025 / 487.7 [0.000, 0.076] / [463.9, 511.5]	0.665 / 337.7 [0.513, 0.817] / [289.3, 386.0]	0.820 / 146.6 [0.690, 0.950] / [106.6, 186.6]

TABLE III: Mode Switching and Pure Surveillance Performance. Each entry reports SR / Time (s), with 95% confidence intervals in brackets.

Variant	SR $\uparrow$	$T_{\mathrm{cap}}$ (s) $\downarrow$	CR $\downarrow$	Var $\downarrow$	P-Steps $\downarrow$
Random	0.000 [0.000, 0.000]	380.9 [303.6, 458.1]	0.1403 [0.1173, 0.1634]	2.392 [2.276, 2.507]	124.8 [113.8, 135.8]
Average	0.750 [0.640, 0.860]	33.0 [30.8, 35.2]	0.0083 [0.0044, 0.0123]	0.836 [0.752, 0.919]	109.4 [85.9, 133.0]
Rule-based	0.400 [0.275, 0.525]	47.6 [43.5, 51.6]	0.0320 [0.0219, 0.0422]	1.056 [0.972, 1.140]	113.6 [105.2, 122.0]
DSGN-WO-BETA	0.656 [0.540, 0.771]	35.4 [33.4, 37.5]	0.0129 [0.0044, 0.0214]	0.756 [0.617, 0.894]	93.3 [74.7, 112.0]
DSGN-SCALAR	0.611 [0.425, 0.797]	34.0 [32.8, 35.2]	0.0148 [0.0074, 0.0222]	0.792 [0.777, 0.807]	96.7 [79.3, 114.2]
DSGN-SCALAR- WO-BETA	0.506 [0.241, 0.770]	34.5 [33.7, 35.2]	0.0223 [0.0095, 0.0351]	0.810 [0.715, 0.905]	125.8 [98.0, 153.7]

TABLE IV: Large-Scale Arbitration Variants (46v12–77v19). Each cell reports the mean with a 95% confidence interval in brackets.

Fig. 5: Temporal coverage dynamics in the 6v2 task-switching scenario.

Scenario	$\mathbf{C_{\text{switch}}}$ $\uparrow$					$\mathbf{T_{\text{resume}}}$ (s) $\downarrow$
Scenario	DSGN	DSGN-WO- BETA	DSGN- SCALAR	DSGN-SCALAR- WO-BETA	TransIQN	DSGN	DSGN-WO- BETA	DSGN- SCALAR	DSGN-SCALAR- WO-BETA	TransIQN
3v1	0.252 [0.216, 0.288]	0.232 [0.169, 0.294]	0.234 [0.217, 0.251]	0.240 [0.205, 0.275]	0.237 [0.201, 0.273]	176.4 [83.1, 269.7]	212.4 [116.3, 308.5]	250.5 [198.8, 302.2]	269.5 [28.0, 511.1]	268.5 [153.8, 383.1]
6v2	0.466 [0.388, 0.544]	0.436 [0.329, 0.543]	0.440 [0.405, 0.475]	0.455 [0.395, 0.515]	0.446 [0.409, 0.483]	55.5 [35.4, 75.7]	69.3 [39.5, 99.0]	89.0 [65.3, 112.7]	100.6 [53.8, 147.5]	99.0 [26.9, 171.1]
9v3	0.636 [0.551, 0.721]	0.581 [0.470, 0.691]	0.610 [0.590, 0.630]	0.620 [0.566, 0.673]	0.608 [0.563, 0.654]	31.6 [20.9, 42.4]	36.3 [23.9, 48.8]	41.4 [36.3, 46.5]	45.9 [33.1, 58.8]	56.9 [17.4, 96.5]

TABLE V: Temporal Indicators for Mode Switching

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 11 — Fig. A.1: Representative 6v2 layouts: G2 (one hexagonal hole, left) and G3 (four triangular holes, right). Green circles labeled P0–P5 are pursuers, magenta circles are obstacles, and red diamonds are surveillance nodes. Gray areas are inaccessible. G0 has no hole.

Method	SR $\uparrow$	$\mathbf{T_{\mathrm{cap}}}$ (s) $\downarrow$	CR $\downarrow$	P-Steps $\downarrow$
TransIQN	0.950 [0.765,0.990]	26.1 [20.8,31.4]	0.0083 [0.0000,0.0258]	227.8 [168.9,286.7]
D-w/o-B	0.935 [0.790,1.000]	30.3 [25.0,35.6]	0.0083 [0.0000,0.0290]	129.6 [109.4,149.8]
D-S	0.965 [0.825,1.000]	28.7 [20.6,36.8]	0.0111 [0.0000,0.0589]	186.3 [86.2,286.3]
D-S-w/o-B	0.885 [0.485,1.000]	30.0 [23.2,36.8]	0.0083 [0.0000,0.0290]	229.3 [186.8,271.7]
DSGN	1.000 [1.000,1.000]	27.1 [18.4,35.8]	0.0000 [0.0000,0.0000]	89.7 [40.5,139.0]

Appendix

TABLE A.II: 6v2 Geometry Sensitivity across DSGN Variants and TransIQN

Scenario	Grid	SR $\uparrow$	$\mathbf{T_{\mathrm{cap}}}$ (s) $\downarrow$	CR $\downarrow$	P-Steps $\downarrow$
3v1	$3\times 3$	0.995 [0.985,1.000]	24.2 [21.8,26.6]	0.0000 [0.0000,0.0000]	218.2 [210.0,226.4]
3v1	$5\times 5$	0.980 [0.961,0.999]	24.5 [21.9,27.1]	0.0083 [0.0000,0.0169]	216.0 [209.1,222.9]
6v2	$3\times 3$	0.970 [0.946,0.994]	33.7 [31.7,35.7]	0.0075 [0.0012,0.0138]	88.8 [84.5,93.1]
6v2	$5\times 5$	0.955 [0.926,0.984]	32.9 [31.2,34.6]	0.0108 [0.0034,0.0182]	86.2 [82.2,90.2]
9v3	$3\times 3$	0.900 [0.858,0.942]	28.8 [27.2,30.4]	0.0128 [0.0072,0.0184]	52.1 [48.6,55.6]
9v3	$5\times 5$	0.880 [0.835,0.925]	27.6 [26.0,29.2]	0.0161 [0.0097,0.0225]	52.9 [49.4,56.4]

Appendix

TABLE A.III: DSGN Action-Granularity Sensitivity in Task Switching

Multi-Robot Coordination

Multi-robot coordination studies how teams of robots plan, communicate, allocate tasks, and act together safely and efficiently.

See all 328 papers

COHORT: Hybrid RL for Collaborative Large DNN Inference on Multi-Robot Systems Under Real-Time Constraints
Mar 11, 2026Mohammad Saeid Anwar, Anuradha Ravi, Indrajeet Ghosh +3Efficient Neural Network Inference Multi-Robot Task Allocation
University of Maryland Baltimore County · DEVCOM Army Research Laboratory
Distributed Multi-Coverage for Robot Swarms
May 20, 2026Mariem Guitouni, Aaron T. BeckerAerial Robotics Multi-Robot Systems
University of Houston, Texas, USA, 77204
SFG-ROS: A Resource-Aware Framework for Dense Multi-Agent Perception
May 22, 2026Constantin Blessing, Elias Geiger, Jakob Häringer +2Multi-Robot Systems Robotics
Institute for Intelligent Systems, Faculty of Computer Sciences and Engineering, Esslingen University of Applied Sciences, Esslingen am Neckar, 73732, Baden-Württemberg, Germany.

Mar 11, 2026J/K move · Enter open · S save

COHORT: Hybrid RL for Collaborative Large DNN Inference on Multi-Robot Systems Under Real-Time Constraints

Mohammad Saeid Anwar, Anuradha Ravi, Indrajeet Ghosh, Gaurav Shinde+2

University of Maryland Baltimore County · DEVCOM Army Research Laboratory

Efficient Neural Network Inference Multi-Robot Task Allocation Multi-Agent Reinforcement Learning

Open paper AskPDF arXiv

Large deep neural networks (DNNs), especially transformer-based and multimodal architectures, are computationally demanding and challenging to deploy on resource-constrained edge platforms like field robots. These challenges intensify in mission-critical scenarios (e.g., disaster response), where robots must collaborate under tight constraints on bandwidth, latency, and battery life, often without infrastructure or server support. To address these limitations, we present COHORT, a collaborative DNN inference and task-execution framework for multi-robot systems built on the Robotic Operating System (ROS). COHORT employs a hybrid offline-online reinforcement learning (RL) strategy to dynamically schedule and distribute DNN module execution across robots. Our key contributions are threefold: (a) Offline RL policy learning combined with Advantage-Weighted Regression (AWR), trained on auction-based task allocation data from heterogeneous DNN workloads across distributed robots, (b) Online policy adaptation via Multi-Agent PPO (MAPPO), initialized from the offline policy and fine-tuned in real time, and (c) comprehensive evaluation of COHORT on vision-language model (VLM) inference tasks such as CLIP and SAM, analyzing scalability with increasing robot/workload and robustness under . We benchmark COHORT against genetic algorithms and multiple RL baselines. Experimental results demonstrate that COHORT reduces battery consumption by 15.4% and increases GPU utilization by 51.67%, while satisfying frame-rate and deadline constraints 2.55 times of the time.

Abstract

Multi-Robot Coordination

Explore similar work

COHORT: Hybrid RL for Collaborative Large DNN Inference on Multi-Robot Systems Under Real-Time Constraints

Distributed Multi-Coverage for Robot Swarms

SFG-ROS: A Resource-Aware Framework for Dense Multi-Agent Perception