Swarm Telescope: Distributed Array Control

• Swarm Telescope is a distributed system that decomposes traditional observatories into autonomous, software-driven elements coordinated via embedded agents.
• It minimizes scheduling delays and human intervention by employing real-time decision-making and dynamic sub-array formation for diverse science cases.
• Implementations like LWA and GOTO demonstrate enhanced uptime, reduced latency, and improved resource efficiency in astronomical observations.

A swarm telescope is an operational paradigm in which a traditionally monolithic astronomical facility is decomposed into a set of smaller, largely autonomous elements that collaborate to function as a single distributed instrument. The approach leverages embedded software-defined agents within each unit, decentralizing control and enabling dynamic, realtime formation of sub-arrays tailored to heterogeneous science drivers, while minimizing human intervention. The swarm telescope paradigm draws from swarm intelligence, emphasizing local autonomy, distributed decision-making, peer-to-peer or organized coordination, and robust, scalable operations—principles that are increasingly essential as observatories grow in complexity, geographic span, and dynamic configurability (Dowell et al., 2018).

1. Conceptual Foundations and Motivations

The swarm telescope concept was formalized by Law et al. in the context of radio interferometers, where array operations are increasingly constrained by software-defined hardware, commensal observing, and large physical footprints (Dowell et al., 2018). The main motivations are:

• Complexity growth: Modern facilities can form multiple simultaneous beams, run parallel projects, or support multi-pointing modes, leading to high configuration and scheduling complexity.
• Geographical distribution: Arrays now often span continents (e.g., VLBA, SKA), complicating monitoring, fault tolerance, and weather-responsive scheduling.
• Scalability and resilience: By dividing the observatory into autonomous elements, operational single points-of-failure are reduced, maintenance is localized, and the system more readily supports incremental growth and robust recovery.

The defining feature is shifting from central, human-driven control to collective decision-making through local agents that monitor, self-optimize, and coordinate via streamlined communication protocols (Dowell et al., 2018). This enables global performance to emerge from local interactions, reducing personnel costs and downtime.

2. System Architecture and Control Layers

A generic swarm telescope comprises three tightly coupled architectural layers (Dowell et al., 2018):

2.1 Autonomous Element Control

Each instrumented element (antenna, dish, optical telescope, or CubeSat) runs a local agent tasked with:

• Monitoring hardware and environmental sensors (e.g., temperature, weather, power, mechanical status).
• Computing operational state via heuristics or ML classifiers (go/no-go logic, anomaly detection).
• Executing autonomous safety and recovery operations (stow, shutdown, power cycling).
• Parsing observational schedules and external triggers, supporting preemption and queue management.

Internally, such agents implement state machines or predictive models mapping telemetry $x(t)$ to operational flags $s \in \{0,1\}$ and recovery actions $R(s)$. Pseudocode examples for main loops and state transitions are provided in (Dowell et al., 2018).

2.2 Inter-Element Communication

Two principal communication models are prevalent:

• Leaderless (broadcast): All elements independently receive external triggers (e.g., GRB, GW alerts) or respond to local transient detections, suitable for small arrays or highly parallel transient searches.
• Organized (hierarchical): One or more organizer agents maintain aggregated global state vectors $S_{global} = \{ s_i, r_i, \theta_i \}_{i=1...N}$, polling element-agents for status and orchestrating resource allocation via constrained optimization. Assignment of elements to tasks involves solving:

$$x^* = \arg\max_x U(x) - \lambda C(x) \quad \text{subject to} \,\, A x \geq b, \, x \in \{0,1\}^N$$

where $U$ is utility (e.g., science priority), $C$ is operational cost, and constraints enforce the required sensitivity/baseline coverage.

2.3 Data Transport and Storage Management

Depending on the array's network and correlator setup:

• Real-time correlators: Local agents monitor data link health, establishing “up” status as a condition for observation.
• Offline or distributed stations: Agents manage local storage pools, schedule data transfers (e.g., via rsync-based workflows), detect file corruption, and coordinate with archival centers without interfering with live observations (Dowell et al., 2018).

3. Scheduling, Optimization, and Swarm Protocols

Swarm telescope scheduling is governed by dynamic, often multi-objective, optimizations executed by organizers or distributed protocols (Dowell et al., 2018):

$$\text{maximize} \quad \sum_j P_j \sum_i x_{i,j} w_{i,j} - \lambda \left(\sum_i c_i x_{i,j} + \text{transition costs}\right)$$

Subject to:

• Minimum elements for each observation: $\sum_i x_{i,j} \geq M_j$
• Exclusive allocation: $\sum_j x_{i,j} \leq 1 \quad \forall i$
• Availability constraints: $s \in \{0,1\}$0 if $s \in \{0,1\}$1

Typical solvers are greedy or message-passing; each element may “bid” for tasks or allocations are centrally orchestrated. Trigger preemption is automatic, reprioritizing queues and reallocating elements without human intervention.

This paradigm supports advanced modes such as commensal observing (multiple projects running simultaneously), dynamic sub-array formation, and opportunistic filler tasks that maximize otherwise idle time.

4. Key Operational Implementations

Several distributed astronomical facilities exemplify the swarm telescope conceptual framework.

4.1 Long Wavelength Array (LWA)

Each LWA station (LWA1, LWA-SV; 256 dual-polarization dipoles) is governed by Python-based element agents (“HAL” systems) that manage both hardware and schedule triggers, with automated data transfer via multi-threaded “SmartCopy.” The system achieves:

• Trigger response times of 30 s–2 min
• Automated protections yielding >98% uptime
• 25% reduction in data loss events, with ∼90% reduction in required human operator intervention (Dowell et al., 2018)

4.2 7-Dimensional Telescope (7DT) and TCSpy

TCSpy provides a three-layer, HTTP/JSON-driven control architecture using ASCOM Alpaca+Alpyca for 20 0.5-m telescopes. Key features include:

• Central scheduling logic achieving ≤1.9 s end-to-end latency for synchronized actions
• Modular “spec,” “deep,” and “search” swarm observation modes supporting spectral mapping, deep co-added imaging, and tiled search/response to gravitational-wave localization maps
• Reliability scaling as $s \in \{0,1\}$2, with rapid reallocation on unit failures (Choi et al., 2024)

4.3 SkyNet Campaign Manager (CM)

SkyNet's Campaign Manager transforms a global network of ∼20 telescopes into a swarm configuration by deploying per-node agents with local queues, regular synchronization to a central dispatcher, and dynamic, credit-normalized scheduling based on per-telescope/filter efficiency and scientific priority. The system supports real-time response to GRB, GW, and SN events, enforces ≥240 s queued time per node to avoid wasted slews, and achieves typical response latencies of ≲ 60 s (Dutton et al., 2022).

4.4 GOTO and Wide-Field Arrays

The GOTO project demonstrates swarm-like scaling with modular units (8 × 40 cm telescopes per mount) coordinated in dual-hemisphere networks, operating in either survey or alert-tracking modes. The scheduler maximizes a utility function $s \in \{0,1\}$3 and supports reconfiguration with 10 s cadence. In O3, the system followed up 52/76 GW alerts within 30–100 s of LIGO–Virgo notifications (Dyer et al., 2020).

4.5 COMCUBE-S Constellation

COMCUBE-S is a constellation of 27 16U CubeSats, each with advanced Compton polarimeters and BGO spectrometers; the network achieves low-latency, multi-unit coincidence triggering for GRBs, collectively providing ≳2 GRB detections per day and a minimum detectable polarization ≤30% for tens of bursts per year (Franel et al., 28 Oct 2025).

4.6 Evryscope

Evryscope employs 23 × 7 cm lenses arranged hemispherically—epitomizing the dense-swarm approach with a 9 060 deg² FOV, 2-min cadence, seamless sidereal tracking, and robust distributed control, resulting in quasi-continuous monitoring of all accessible sky (Law et al., 2014).

5. Quantitative Improvements over Traditional Models

Operational metrics consistently demonstrate that swarm telescopes outperform classical arrays on multiple axes (Dowell et al., 2018, Law et al., 2014, Dyer et al., 2020, Choi et al., 2024, Dutton et al., 2022):

• Reduced latency: Scheduling delays shrink from 10–60 min to ≲ 30 s (automatic); recovery from failure, from hours to <5 min.
• Uptime and resilience: >98% automated protection in the LWA; negligible “dead” time, as autonomous shutdowns prevent cascading faults and organizers reallocate resources dynamically.
• On-sky time and cost: 15–20% more productive observation time, 30% lower operational budget observed in the LWA case.
• Resource efficiency: Automated selection of sub-arrays and dynamic project filling increases array utilization by up to 10% per element.

A summary table from the LWA deployment (Dowell et al., 2018):

Metric	Classical Array	Swarm Telescope
Scheduling delay	10–60 min (manual)	~30 s (automatic)
Recovery from hardware/weather failure	2–4 hr	<5 min
Operational cost (relative)	100%	~70%
Uptime	80–90%	>98%
Human ops involvement	100%	~10%

6. Advanced Observing Modes and Scheduling Models

Swarm telescope architectures enable a range of advanced operating modes:

• Dynamic sub-arraying: Organizers can cluster elements as phased arrays for VLBI or core pulsar timing, or allocate elements based on instantaneous weather, sensitivity, or commensal science demands (Dowell et al., 2018).
• Commensal/filler observing: Idle units are dynamically assigned lower-priority projects, boosting aggregate utilization (Dowell et al., 2018, Choi et al., 2024, Dutton et al., 2022).
• Trigger-driven scheduling: Real-time preemption of ongoing observations in response to high-priority transients (GRBs, GW events), with reallocation and resource trade-offs computed in seconds (Dyer et al., 2020, Choi et al., 2024, Dutton et al., 2022).
• Coordinated tiling: Arrays like GOTO and 7DT participate in rapid, distributed tiling of localization regions for transient follow-up, synchronized within <1 s jitter (Dyer et al., 2020, Choi et al., 2024).
• Global optimization constraints: Sensitivity and sky coverage requirements are enforced mathematically, e.g., $s \in \{0,1\}$4, with costs in slew, power, and other resources (Dowell et al., 2018).

7. Scalability, Extensibility, and Future Directions

The swarm paradigm is directly extensible to arrays with $s \in \{0,1\}$5 elements, supported by modular API design (e.g., ASCOM Alpaca in TCSpy (Choi et al., 2024)), scalable thread pools, and hierarchical organization into regional controllers or organizers. Critical lessons from existing deployments include the necessity for robust, local control to prevent cascading failures and the need for rigorous cross-element calibration methods (flat-fielding, systematics management) (Law et al., 2014).

Proposed extensions to next-generation facilities, such as the ngVLA (214 × 18-m dishes) envisage multi-tier organizer structures managing geographically clustered sub-arrays, real-time resource allocation for wide-ranging science cases, and integrated self-triggering from embedded transient pipelines (Dowell et al., 2018). In spaceborne contexts, the COMCUBE-S design emphasizes rapid inter-satellite communication, systematics monitoring, and optimized sky coverage from equatorial LEO (Franel et al., 28 Oct 2025).

Collectively, these developments underscore that the swarm telescope model offers a scalable, fault-tolerant, and scientifically agile operational approach for the next era of large-scale, distributed observatories across the electromagnetic spectrum.