Source-linked AI summary
DSGN: A Dual-Stream Gated Network for Scalable Simultaneous Surveillance and Containment
Heng Zhang, Ziyang Zuo, Guoxiang Zhao, Xiaoqiang Ren
TL;DR
SSC requires robots to balance dispersed area coverage with cohesive target enclosure, while conventional shared representations may mix those competing cues. DSGN uses separate task streams coordinated by a context-aware gate with a Beta prior; evaluations report strong task outcomes across scales, including top or tied-top scale-averaged task completion and lowest scale-averaged collision rate.
Problem
SSC couples persistent area surveillance with post-detection containment, while conventional unified policies can mix dispersion-oriented coverage and target-centric enclosure cues.
Source: §I · pp. 1–2Method
DSGN processes observations in separate surveillance and containment streams, fuses them through context-aware gating, and regularizes the gate with a state-dependent Beta prior.
Source: §I · p. 2, B. Network Architecture · p. 3, D. Training Strategy via Probabilistic… · p. 4Results
DSGN achieves the highest or tied-highest scale-averaged Strict SR and the lowest scale-averaged CR at all three scales, with the fastest variant’s Tcap only 0.1–2.1 s lower.
Source: D. Results and Discussion · p. 6Takeaways & Limitations
Across the reported evaluations, task-specific streams and gating support safer, more consistent coordination as interacting-agent counts increase and post-containment surveillance recovers across team sizes.
Source: D. Results and Discussion · p. 6, §V · p. 8
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
SSC couples persistent area surveillance with post-detection containment, requiring robots to balance coverage and target enclosure. DSGN separates these objectives into specialized streams and uses context-aware gating with a Beta prior to coordinate task transitions.
I. INTRODUCTION
- SSC requires multi-robot systems to balance continuous area coverage with timely containment of targets after detection.Surveillance and containment have often been studied separately, but SSC couples them in one problem.Source: §I · p. 1
- Prior work addresses search or containment separately, leaving practical SSC settings in which targets are initially unknown to require both capabilities.Distributed containment approaches include differential games, bio-inspired rules, and DRL; surveillance work includes coverage, localization, and target search.Source: §I · p. 1
- Analytical switching is difficult because desired behavior depends jointly on coverage, target observations, robot distribution, and containment geometry.Fixed or hand-crafted gates may adapt poorly, while naive adaptive gating may chatter and cause task failures.Source: §I · p. 2
- DSGN separates surveillance and containment into task-specific streams and coordinates them with context-aware gating regularized by a Beta prior.The design targets stable transitions when target observations are intermittent or task demands change.Source: §I · p. 2
A. Dec-POMDP Formulation and System Abstraction
- The system is formulated as a decentralized partially observable Markov decision process, with pursuers acting from local observations of nearby entities and obstacles.The mission combines persistent surveillance of the region with containment of all evaders.Source: A. Dec-POMDP Formulation and System… · p. 2, §II · p. 2
B. Kinematic Objectives of the SSC Problem
SSC combines two kinematically different objectives: pursuers disperse to reduce surveillance-node idleness and contract around evaders for containment. Its mission objective is to keep idleness low while satisfying the containment condition for every evader.
B. Kinematic Objectives of the SSC Problem
- Surveillance requires pursuers to disperse across the region to reduce node idleness, while containment requires them to contract around evaders.The two objectives therefore call for different spatial arrangements of the same team.Source: B. Kinematic Objectives of the SSC… · pp. 2–3
- The mission objective is to maintain low surveillance idleness while satisfying the containment condition for all evaders.The containment condition is described as ensuring closure of the escape cone and stability of the formation.Source: B. Kinematic Objectives of the SSC… · p. 3
III. APPROACH
DSGN uses separate spatial-explorative and tactical-exploitative reasoning streams, coordinated by a context-aware gate, to produce a unified decentralized policy. Its pipeline encodes structured observations, models relations, and fuses task-specific contexts.
A. Observation and Input Representation
- Each pursuer’s structured input includes ego, teammate, evader, obstacle, and surveillance-node information, with entity-specific features derived from local observations.Features include relative velocity, engagement intent, heading-orientation error, obstacle radius, and normalized node idleness.Source: A. Observation and Input Representation · p. 3
B. Network Architecture
- DSGN maps structured inputs to discrete kinematic commands through feature projection, relational encoding, task-disentangled fusion, and distributional action inference.The architecture aligns entity types in a common latent space, reasons about spatial relations, then arbitrates task-specific contexts.Source: B. Network Architecture · p. 3, Figure 2 · p. 3
1) Heterogeneous Feature Projection:
- Type-specific multilayer perceptrons project heterogeneous tokens into a common latent space, while type embeddings preserve categorical identity.A validity mask prevents padded entities from contributing to attention.Source: 1) Heterogeneous Feature Projection · p. 3
2) Relational Context Encoding:
- A transformer encoder models high-order dependencies among projected entities and aggregates scene geometry, teammate distribution, evader observations, obstacles, and surveillance urgency.The updated ego feature serves as the central query for dual-stream decision-making.Source: 2) Relational Context Encoding · pp. 3–4
3) Task-Disentangled Gated Fusion:
DSGN maintains distinct surveillance and containment contexts, then uses a vector-wise gate to arbitrate between them. It maps the fused representation to actions using an implicit quantile network that estimates action-return distributions.
3) Task-Disentangled Gated Fusion:
- Surveillance attention highlights neglected regions for coverage, while containment attention represents target geometry, teammate support, and local formation needs.The containment stream still processes teammate configurations when no evader is observed, preserving tactical readiness without forcing premature containment.Source: 3) Task-Disentangled Gated Fusion · p. 4
4) Distributional Action Inference:
- Large vector-gate values emphasize containment channels and small values emphasize surveillance channels, allowing gradual commitment as target evidence becomes reliable.IQN maps the fused representation to actions by approximating each action’s return distribution, useful when outcomes vary across partially observed local contexts.Source: 3) Task-Disentangled Gated Fusion · p. 4, 4) Distributional Action Inference · p. 4
C. Multi-Objective Reward Design
The shaped reward converts the SSC objective into dense signals for decentralized policy optimization, combining surveillance, containment, safety, and efficiency requirements.
C. Multi-Objective Reward Design
- The shaped reward translates the task-level SSC objective into dense learning signals for decentralized policy optimization.It covers surveillance, containment, safety, and efficiency.Source: C. Multi-Objective Reward Design · p. 4
- A potential-based term directs pursuers toward the surveillance node with the greatest idleness, providing a smooth gradient toward neglected regions.The shaping follows policy-invariant reward shaping.Source: C. Multi-Objective Reward Design · p. 4
- Containment guidance limits each evader’s engagement set to the Neng closest pursuers, reducing overcrowding around the same target.It uses the distance to the closest observed evader to determine engagement and guidance.Source: C. Multi-Objective Reward Design · p. 4
D. Training Strategy via Probabilistic Prior Injection
The training strategy regularizes DSGN’s vector gate with a state-dependent Beta prior and uses a progressive curriculum to increase the number of entities during training.
D. Training Strategy via Probabilistic Prior Injection
- A state-dependent Beta prior regularizes each gate channel, favoring low containment activation without observed evaders and decisive gating when evaders are visible.The prior uses (α, β) = (1, 4) when no evader is visible and (0.5, 0.5) when one is visible.Source: D. Training Strategy via Probabilistic… · p. 4
A. Training Setup
- The simulator adds surveillance nodes, reactive evaders, and multi-target containment, while a progressive curriculum gradually increases entity counts.Agent inputs include at most 5 obstacles, 5 teammates, 8 evaders, and 16 surveillance nodes; unused entries are zero-padded.Source: A. Training Setup · p. 5
B. Evaluation Setup
The evaluation tests structural generalization across constant-density scales and compares DSGN with adapted decentralized, motion-planning, reinforcement-learning, and heuristic baselines. DSGN performs strongly on scale-averaged task completion, safety, and post-containment coverage metrics.
B. Evaluation Setup
- Policies are tested without retraining across nine constant-density configurations from 12v3 to 77v19, with 40 trials per seed and five independent seeds.The largest evaluation setting, 77v19, exceeds the largest curriculum training setting, 15v4.Source: B. Evaluation Setup · p. 5, Table II · p. 5
C. Baseline Methods
- The baselines are TransIQN, IQN, DQN, and a reactive APF heuristic; TransIQN shares DSGN’s transformer backbone but uses a single ego-token feature.Learned methods share the shaped reward, curriculum, action space, and 7M-step training budget per seed.Source: C. Baseline Methods · pp. 5–6
D. Results and Discussion
- Across scales, DSGN has highest or tied-highest scale-averaged SR, lowest CR and P-Steps, and Tcap only 0.1–2.1 s above the fastest variant.The best variant can vary by configuration, and DSGN does not lead in every individual configuration.Source: Figure 4 · p. 6, D. Results and Discussion · p. 6
E. Ablation Study
- The ablations compare random, equal-weight, rule-based, no-Beta, scalar-gate, and scalar-gate-without-Beta arbitration variants.The main comparisons include DSGN and its gate and prior ablations across the nine scale-density configurations.Source: E. Ablation Study · p. 6, D. Results and Discussion · p. 6
1) Dynamic Gating and Probabilistic Priors:
Removing either vector-wise gating or the Beta prior worsens several scale-averaged metrics. Phase-specific tests, geometry and action-granularity evaluations, and the conclusion characterize DSGN’s switching performance and scope.
1) Dynamic Gating and Probabilistic Priors:
- Removing the Beta prior or vector-wise gating worsens scale-averaged SR, CR, Var, and P-Steps at all three scales, while Average retains a modest Tcap advantage.The complete scale comparison is in Table A.I; Table IV shows only large-scale results.Source: 1) Dynamic Gating and Probabilistic… · p. 6, Table IV · p. 7
2) Mode Switching Across Task Phases:
- In pure surveillance and task switching, DSGN has Pure Surveillance SR of 1.000 at every tested team size and the shortest mean task-switching duration in all settings.With scalar gating, adding the Beta prior improves Task Switching SR and shortens mean episode duration in every configuration, while also raising Pure Surveillance SR at every tested team size.Source: 2) Mode Switching Across Task Phases · pp. 6–7, Table III · p. 7
3) Sensitivity to Search-Region Geometry:
- Across 6v2 layouts, DSGN has the highest mean SR and lowest mean P-Steps, but its SR falls from 1.000 with no holes to 0.500 with four triangular holes.The two hole layouts each block 20.0% of the workspace, testing spatial arrangement at equal blocked area.Source: 3) Sensitivity to Search-Region Geometry · p. 7
4) Sensitivity to Action Granularity:
- With 5 × 5 commands, DSGN remains effective and maintains comparable execution efficiency, although SR is lower and CR higher than with 3 × 3 commands in all three settings.The finer grid has mean Tcap and P-Steps close to those of 3 × 3.Source: 4) Sensitivity to Action Granularity · pp. 7–8
V. CONCLUSION
- DSGN uses vector-wise gating with a state-dependent Beta prior; future work will examine continuous actions, repeated target appearances, and 3D nonholonomic robots with limited communication.Source: §V · p. 8
APPENDIX A ADDITIONAL EXPERIMENTAL RESULTS
Appendix A reports arbitration results across three network scales, geometry sensitivity across hole layouts, and task-switching sensitivity to action granularity.
APPENDIX A ADDITIONAL EXPERIMENTAL RESULTS
- For task switching across the 5×5 action grid, Average has SR 0.917 and Rule-based has SR 0.800.The grid varies acceleration and angular velocity over the values specified in the table note.Source: Appendix A · pp. 9–10
- DSGN records SR 0.922 at small scale, 0.811 at medium scale, and 0.750 at large scale.The listed settings are 12v3–20v5, 38v10–64v16, and 46v12–77v19, respectively.Source: Appendix A · p. 9
- At medium scale, DSGN-WO-BETA has SR 0.711, DSGN-SCALAR has 0.750, and DSGN has 0.811.The medium-scale setting is 38v10–64v16.Source: Appendix A · p. 9
- In the G2 layout with one hexagonal hole, TransIQN has SR 0.650, versus 0.835 for D-w/o-B and 0.765 for D-S-w/o-B.Brackets in the table denote 95% confidence intervals; the variants are identified in the table note.Source: Appendix A · p. 10
- In the G3 layout with four triangular holes, TransIQN has SR 0.350, versus 0.485 for D-w/o-B and 0.450 for D-S-w/o-B.Random has SR 0.433 in this layout.Source: Appendix A · pp. 9–10