Swarm Multi-Target Tracking: DSBM+ASMA

• Swarm Multi-Target Tracking (DSBM+ASMA) is a decentralized approach where autonomous agents coordinate via adaptive behavior models and distributed memory to effectively track dynamic targets.
• The DSBM component utilizes attractive and repulsive force-based cues to balance exploration and exploitation, enhancing coverage and persistence in complex environments.
• ASMA underpins robust information sharing with multi-hypothesis assignment and dynamic split-merge strategies, enabling adaptive agent reallocation in response to real-time target dynamics.

Swarm Multi-Target Tracking (DSBM+ASMA) encompasses a paradigm of decentralized target tracking in which collections of autonomous agents coordinate via locally computable behavioral rules and distributed memory or hypothesis structures. The Dynamic Social Behavior Model (DSBM) provides adaptive, situation-aware movement or behavioral priors for each agent, while the Agent-based Social Memory Architecture (ASMA) encompasses distributed memory, multi-hypothesis assignment, or partitioning/split-merge strategies. Such frameworks enable robust, scalable tracking of multiple dynamic targets, often in adversarial or resource-constrained environments, across modalities from ground robots to aerial and spaceborne swarms.

1. Problem Definition and Context

Swarm multi-target tracking addresses the challenge of collectively searching for, localizing, and persistently tracking multiple dynamic targets using a distributed group of agents without a centralized coordinator. Scenarios include (i) ground or aerial robots seeking to track fast or evasive objects, (ii) drones operating in complex 3D environments with obstacles and partial observability, and (iii) spacecraft performing angles-only tracking in resident space object surveillance. Constraints typically include local sensing (often binary or bearing-only), limited inter-agent communication (restricted to k-nearest neighbor graphs or finite radii), and real-time operation without global knowledge or central control (Kwa et al., 2021, Qamar et al., 2022, Kruger et al., 2020).

A central objective is to maximize time-averaged coverage (the fraction of time any agent senses a target within range) or measurement assignment precision, while dynamically balancing between exploration (searching the environment) and exploitation (converging on known or predicted target locations). DSBM+ASMA frameworks have been concretely motivated by the difficulty of tracking targets that move faster than the agents themselves, or in the presence of severe observational aliasing (e.g., in angles-only tracking where absolute positions are unobservable).

2. Dynamic Social Behavior Model (DSBM)

DSBM is a formalism for generating adaptive agent behaviors—whether physical motion or data association—using a composition of locally computable forces, priors, or kinematic models.

2.1 Decentralized Force-Based Models

For ground and aerial swarms, DSBM typically superimposes an attractive (exploitative) term that draws agents towards recent target locations, and a repulsive (exploratory) term that disperses agents to improve area coverage. Specifically, each agent updates its velocity vector according to

$v_i(t) = v_{i,att}(t) + v_{i,rep}(t)$

where $v_{i,att}$ is computed from the agent’s or its neighbors’ most recent target sighting (via ASMA), and $v_{i,rep}$ sums repulsive social forces from the $k$ nearest neighbors, scaled by an adaptively tuned coefficient $a_R$ (Kwa et al., 2021). The attractive term is modulated by exploitation bias and stochasticity, while the repulsive term is adaptively increased for agents not currently tracking a target (search mode) and decreased for agents in tracking mode.

2.2 Domain-Specific Modeling in Non-Sensing Modalities

In angles-only tracking of space objects, the DSBM formalism is redeployed as a measurement transformation that produces a parametric model for target bearing evolution in the observer-centric frame. By using differential (relative) angle measurements between simultaneously viewed targets, perturbative distortions such as $J_2$ or drag are largely canceled, enabling low-order curve models for azimuth/elevation evolution as a function of orbital true anomaly (Kruger et al., 2020).

3. Agent-based Social Memory Architecture (ASMA)

ASMA unifies several strategies for distributed storage and propagation of information underpinning exploitative actions or multi-hypothesis association.

3.1 Peer-Shared, Time-Windowed Memory

In robotic swarms, each agent maintains a local memory of (i) the last detected target position and (ii) timestamp. At each timestep, agents exchange their memory entries with their $k$ nearest neighbors, and select the most recent (valid) detection as their “point of attraction.” Memory entries have a finite validity window $t_{mem}$: expired entries are cleared, balancing the risk of exploiting stale information against the need for exploitation continuity (Kwa et al., 2021).

3.2 Assignment-Split and Split-Merge Architectures

In tracking-by-detection contexts without direct observations (e.g., optical angles-only), ASMA encompasses assignment-split multi-hypothesis schemes. Agents maintain multiple data association hypotheses, each scored according to residuals against the DSBM-derived kinematic model. Hypothesis selection, pruning, and confirmation are implemented via gating, clustering (e.g., DBSCAN for unassigned points), and statistical scoring, ensuring robustness to ambiguous measurements and associational multiplicity (Kruger et al., 2020). In deep RL-based aerial swarms, ASMA governs dynamic partitioning of the swarm into sub-swarms (island modeling) tracking individual targets, managing autonomous split/merge and reassignment as the number or configuration of targets evolves (Qamar et al., 2022).

4. Exploration–Exploitation Regulation and Distributed Decision Dynamics

The interplay between exploration (rapid reacquisition, environment search) and exploitation (aggregation, tight tracking) is dynamically regulated by tuning DSBM and ASMA parameters. Key controls include:

• Network degree $k$: Larger $k$ increases social information sharing (exploitation), but reduces independent exploration; optimal $k$ shifts lower as the number of targets increases (Kwa et al., 2021).
• Memory length $t_{mem}$: Longer $t_{mem}$ prolongs exploitation, but increases susceptibility to stale bias; optimal values depend on target dynamics and revisit times.
• Adaptive repulsion $a_{R,i}(t)$: Short-term switching between exploration and exploitation is modulated online by an agent’s tracking state (tracked/not-tracked).
• Split/Merge thresholds: For partition-capable swarms, ASMA manages agent allocation to match the current number of targets, ensuring balanced distribution and recombination as situations evolve (Qamar et al., 2022).

This decentralized feedback loop admits parameter tuning to achieve an optimal trade-off between coverage and persistence for any given agent/target speed ratio, number of targets, or agent capabilities.

5. Algorithmic Implementation and Decision Loops

A generic decision loop for DSBM+ASMA frameworks in robotic settings involves the following steps per agent and timestep (Kwa et al., 2021):

1. Sense local environment for target presence within range.
2. Update local memory and encode new target detections.
3. Exchange memory/state with $k$ nearest neighbors.
4. Select point-of-attraction according to most recent detection.
5. Update tracking/exploration state.
6. Adapt repulsion/exploration parameter as a function of current state.
7. Compute attractive and repulsive velocities.
8. Integrate motion, respecting movement constraints (e.g., max velocity).

In assignment-based scenarios, the loop comprises: receiving new measurements, applying DSBM transform to existing tracks, curve-fitting to predict next measurement location, gating by kinematic and behavioral criteria, assignment/branch creation, multi-hypothesis scoring and pruning, and unambiguous measurement confirmation (Kruger et al., 2020).

6. Experimental Results, Performance, and Analysis

DSBM+ASMA frameworks have been validated against challenging benchmarks.

6.1 Robotic and Simulated Swarms

In simulated swarms of $N=50$ ground robots tracking $M=1$–$3$ targets (including evasive targets with $v_o,v_{max}=2v_a,max$), coverage and engagement ratios were analyzed over sweeps of $k$, $t_{mem}$, and speed ratios. Peak tracking performance occurs at intermediate $k^*=15$–$20$ for $M=1$; introduction of moderate memory ($t_{mem}\approx20$ timesteps) is critical for successful tracking of fast targets, with coverage dropping to zero without memory. Increasing $t_{mem}$ beyond a threshold yields over-exploitation and steady performance loss; as $M$ increases, less exploitation (lower $k$) is optimal (Kwa et al., 2021).

In physical trials with $N=6$ robots and a single fast (evasive) target, mean coverage improved from $37.2\%\pm4.6$ (no memory) to $49.4\%\pm9.2$ ($t_{mem}$ enabled), with only memory-enabled swarms maintaining persistent aggregation on the last-seen position.

6.2 RL-based Drone Swarms

Swarms of up to 18 agents trained via Soft Actor–Critic (SAC), PPO, or TPPO in 3D obstacle-rich environments achieved highest mean cumulative rewards with SAC ($\approx1500$), with real-time multi-target split/merge, narrow passage traversal, and robust multi-scenario performance (Qamar et al., 2022).

6.3 Angles-Only Tracking in Space

In the SAMUS system, measurement assignment precision exceeded $99.7\%\pm1.8$ with recall $96.3\%\pm6.0$, average assignment error $\approx27''$, and typical runtimes per epoch of $18\pm6$ ms for $T=3, M\approx10, n=5$. Performance persisted under high measurement noise, occultation, accelerative maneuvers, and orbit errors, substantially outpacing GNN, JPDA, and PHD filters (Kruger et al., 2020).

7. Practical Guidelines and Comparative Insights

• Agent-based memory (ASMA) is essential for tracking targets moving faster than the agents; $t_{mem}$ should be chosen based on target revisit rates to avoid staleness.
• Excessively large social network degree $k$ or memory length $t_{mem}$ induce over-exploitation, with agents failing to disperse and reacquire fast-moving or multiple targets.
• Partitioned swarm architectures (split/merge) are effective in managing agent distribution for dynamic multi-target scenarios (Qamar et al., 2022).
• DSBM+ASMA can be adapted variously as direct physical control laws, distributed memory/split-merge logic, or multi-hypothesis assignment frameworks as dictated by sensing and operational constraints.
• For increasing swarm size $N$, robustness to stochasticity improves, but the need for precise parameter tuning becomes more acute; optimal exploration/exploitation balance sharpens.

In sum, DSBM+ASMA frameworks are a foundational strategy for achieving robust, decentralized, and high-fidelity multi-target tracking in both robotic and sensing-limited domains, with proven effectiveness across terrestrial, aerial, and orbital modalities (Kwa et al., 2021, Qamar et al., 2022, Kruger et al., 2020).