A Comprehensive Review of Shepherding as a Bio-inspired Swarm-Robotics Guidance Approach

I. INTRODUCTION

• Decentralized control lets agents process and act independently, supporting simpler robots and system robustness, flexibility, and scalability,.Source: §I · p. 1
• Shepherding assigns high-level path planning and task allocation to a shepherd, while swarm members handle their own single-agent dynamics.The review frames this as a way to separate the two control functions in a complex multi-agent problem.Source: §I · p. 1
• The review covers shepherding research to assess its potential as a simple, effective control approach for varied swarm-robotic systems.Source: §I · pp. 1–2

II. SHEEPDOG TRIALS

• The International Sheep Dog Society’s trial rules decompose shepherding into basic skills and offer scenarios and scores that may inform benchmarks and AI reward functions.Source: §II · p. 2
• Trials pair a handler’s higher-order decisions with a sheepdog’s basic autonomy, coordinating commands and movement to manage pressure on the sheep.Source: §II · pp. 2–3

III. SHEPHERDING FOUNDATIONS

• The Robot Sheepdog Project used a camera-equipped robot to group ducks and repel them toward a goal, and later work guided real ducks there –.Source: §III · p. 3
• After 250 generations of simulation-based genetic-rule development, a robotic shepherd guided another robot to a goal in 67% of trials.Source: §III · p. 3
• A T-Rex herding simulated Raptors demonstrated objective-based behavior modeling for computer animation.Source: §III · p. 3
• These early studies combined simulation and robot experiments with AI techniques and applied shepherding to varied swarm-control problems –.Source: §III · p. 3

IV. CLASSIC SHEPHERDING

• Two shepherds consistently outperformed one at guiding the swarm under Miki and Nakamura’s rules for cohesion, separation, escape, and shepherd cooperation.Source: §IV · pp. 3–4
• Str¨ombom et al.’s single-shepherd heuristic uses local-neighborhood attraction and driving and collecting behaviors, but its less strategic repositioning may increase movement and energy use.Source: §IV · p. 4
• Hoshi et al. found that faster shepherd motion improved performance against equal-speed swarms, while large speed disparities among variable-speed agents hindered herding.Source: §IV · p. 4
• In 3D simulations, Hoshi et al.’s shepherd guided a wider range of swarm speeds than in 2D and performed similarly with variable-speed swarms.Source: §IV · p. 4
• Tsunoda, Sueoka and Osuka found that an 8° standard deviation in shepherd-to-swarm repulsion error impaired guidance, while increased shepherd speed improved performance.Source: §IV · p. 4
• Zig-zag motion emerged or was prescribed as a guidance strategy across several shepherding studies –,,,,.Source: §IV · p. 4

V. SHEPHERDING CONTROL METHODS

• Force-vector control adapts flocking behaviors including collision avoidance, velocity matching, and centering, with shepherd repulsion used as a key control mechanism,.The Str¨ombom et al. schematic depicts shepherd repulsion and attraction toward the sheep center of mass alongside the sheep’s total velocity vector.Source: §V · p. 4, Figure 1 · p. 5
• Roadmap-based shepherding with 30 swarm agents converged faster than A* path planning, though an extended control method did not perform well generically [49],.Source: §V · p. 5
• Modeling a swarm as a deformable shape and steering it through points behind the swarm outperformed the authors’ earlier shepherding methods.Source: §V · p. 5
• Immune-system-inspired control was ineffective in one study; a follow-up improved results but only tested swarms of two to four agents and could not contain all agents in the target region,.Source: §V · p. 5
• Bennett and Trafankowski’s rule-based method produced good results but was sensitive to its switching angle; recreated baselines may differ from their original implementations [24].The comparison found Vaughan et al.’s method performed poorly in an open environment, while methods by Lien et al. and Miki and Nakamura were generally effective.Source: §V · p. 6
• Trajectory optimization with non-holonomic constraints produced a practical path for one wheeled robot pair but did not extend to controlling an entire swarm.Source: §V · p. 6
• V-formation control achieved higher task success rates but generally took longer than Str¨ombom et al.’s method; two shepherds guided the swarm in fewer time steps than one,.The original V-formation also outperformed the shift V-formation technique.Source: §V · p. 6
• With local-camera positions, furthest-agent targeting always guided the swarm to the goal, while the method derived from Str¨ombom et al. mostly failed.Center-targeting performed similarly with global and local perspectives.Source: §V · p. 6

VI. ALTERNATIVE SHEPHERDING BEHAVIORS

• More shepherds increased guidance speed, with the largest improvement between one and two shepherds, except intensified swarm separation shifted it to two and three.Lien et al. also described herding, covering, patrolling, and collecting as applicable shepherding behaviors.Source: §VI · pp. 6–7
• SOCS combines collecting and herding; its heuristic handled online shepherding through complex obstacles, but the decision model often failed the time limit.The model did not effectively trade off completion time against the number of agents collected and shepherded to the goal.Source: §VI · p. 7
• Cimler et al. modeled ancient Celtic animal herds over about 70 km² and 40 years, with daily time steps and a shepherd leading the swarm.The review says shorter timelines would be needed to assess robotic applications, unless real-time swarm dynamics were negligible.Source: §VI · p. 7

A. Robotic Shepherding

• Splitting and merging UGV groups, each assigned a UAV shepherd, helped agents traverse complex terrain while exploring unknown areas.The approach combined airborne and ground-camera views, though its methods and results discussion were limited.Source: A. Robotic Shepherding · p. 7
• Connected-components labeling recognized independent robotic flocks with accuracy above 70%, but setup details and performance in more complex systems remained unclear.The review notes little information about the simulated setup.Source: A. Robotic Shepherding · p. 7
• With no computation or memory, simulated e-puck shepherds used line-of-sight sensing and evolved reactive control, often circling agents before moving them toward the goal.Swarm agents sensed only shepherds and other agents.Source: A. Robotic Shepherding · p. 7
• Adding Hilbert space-filling-curve behavior made simulated shepherding more efficient at reaching all goals and more effective at reaching a higher goal density.The study also used continuous distance-modulated forces and circular shepherd path planning.Source: A. Robotic Shepherding · p. 7

B. Animal Shepherding

• After two further trials, sheep’s initial 60 m repulsion distance from SCRUPAL had fallen to 10 m, but the single-robot study supported no general conclusions [11].The robot shepherded 30 sheep on a 2.5-hectare paddock.Source: B. Animal Shepherding · p. 8
• Pololu m3pi robots replicating sheepdogs guided a swarm agent to a goal despite experimental uncertainties and errors.The work discussed possible UAV use for large-scale cattle herding, motivated by helicopter-related casualty rates.Source: B. Animal Shepherding · p. 8
• Bat-Erdene and Mandakh’s perimeter-based livestock framework could enclose and move livestock between pastures, but it had no experimental or simulation validation..Source: B. Animal Shepherding · p. 8
• The n-wavefront algorithm diverted birds from airport airspace better than center-of-mass guidance, while flock shape was better maintained in 3D than 2D.A later m-waypoint algorithm was effective in shepherding egret and loon flocks away from protected airspace in experiments and simulations.Source: B. Animal Shepherding · p. 8

VIII. HUMAN-SHEPHERDING INTERACTIONS

• Human-machine shepherding with laser-pointer input and real-time roadmap updates generally outperformed only-human or only-computer simulations on completion time.A projected simulation was captured by camera for user interaction.Source: §VIII · p. 8
• In shepherding games, human pairs commonly used search-and-rescue followed by coupled-oscillatory-containment, while virtual-reality players adapted their behavior as agent speed increased,.An expert artificial agent convinced 12 of 18 players they were playing another human; a separate study found some players discovered a slower oscillatory alternative without a partner,.Source: §VIII · pp. 8–9
• In simulations and actual experiments, a human-operated MITRO telepresence robot shepherd augmented insect-inspired swarm foraging versus a shepherdless honeybee-based swarm, particularly as task complexity increased.Source: §VIII · p. 9

IX. MACHINE LEARNING FOR SMART SHEPHERDING

• Gee and Abbass used machine education to teach supervised neural networks shepherding by dividing it into driving and collecting sub-skills and training each with human demonstrations.The review describes machine learning and computational intelligence for shepherding as still in its infancy, with a large search space making generalization difficult.Source: §IX · p. 9
• Apprenticeship Bootstrapping aggregates models learned from demonstrated sub-skills into an overall shepherding UAV.Earlier work applied the approach to UAVs following ground-vehicle swarms while keeping them in view,.Source: §IX · p. 10
• A genetic algorithm optimized shepherd-force modulation to save energy, though the fixed shepherding model constrained the optimization..Source: §IX · p. 10
• Watchdog AI proposed shepherd agents that monitor other AI and apply corrections, but the framework was presented theoretically without a case study..Source: §IX · p. 10

• Most reviewed work includes simulation; studies with path planning beyond force-vector reactions form the largest category, while patrolling follows herding as an objective.UAVs are the most common UxVs used for shepherding tasks.Source: §X · p. 10
• The review groups remaining challenges into modeling, machine learning and AI, robotics, and shepherding analytics.Source: §X · pp. 10–11
• Figure 3 groups literature by simulation, robotics, AI, or combinations, then divides shepherding taxonomy into guidance and implementation; Table I serves as a glossary and trend reference.The figure expands the relationships among shepherding elements for finding specialized references.Source: §X · p. 10, Figure 3 · p. 11

A. Modelling

• Human-designed swarming models can reflect perception biases because mathematically identified swarming regions may escape human detection.The review questions whether models based on human perception are biased or incomplete.Source: A. Modelling · p. 11
• Sheepdog-inspired models are not guaranteed to be efficient for robots, whose memory, processing, energy, and physical constraints differ.Transferring sheepdog behavior to robots does not ensure equivalent effectiveness or efficiency.Source: A. Modelling · pp. 11–12
• Reviewed models are difficult to adapt to dynamic environments and usually do not generalize to other shepherding forms.The review notes effort in transferring models across contexts and effects from parameter changes such as swarm size.Source: A. Modelling · p. 12
• Machine teaching depends on human experts to provide demonstrations, but experts are costly and time-limited, and emerging shepherding tasks lack standards.These constraints limit collecting sufficiently diverse and numerous demonstrations.Source: A. Modelling · p. 12

B. Intelligence

• The review describes agent-discovered shepherding rules as an early research direction with opportunities for computational intelligence and machine learning.Source: B. Intelligence · p. 12
• Learning swarm-agent behavior could help a shepherd adjust influence-vector characteristics and optimize its behavior in real time as context changes.The review notes behavioral variation within and across sheep breeds.Source: B. Intelligence · p. 12
• Existing models use weighted sums of forces, but their weights depend on swarm and shepherd numbers, terrain, sensor information, and uncertainty.The review calls for AI that can plan and adapt shepherd behavior in dynamic, uncertain environments.Source: B. Intelligence · p. 12
• Activity recognition remains an open issue: a shepherd may need to respond to herd activities, but little research addresses activity recognition in animals.Source: B. Intelligence · p. 12

C. Robotics

• With one shepherd, high-level planning remains centralized, so shepherd malfunction can cause system failure; multiple independently operating shepherds can mitigate this risk.Source: C. Robotics · p. 12
• A shepherd may keep trying to collect a failed agent, preventing the swarm from reaching its destination; ignoring agents that have not moved for a set number of steps is one simple proposed response.The review notes that more complex solutions may be needed.Source: C. Robotics · p. 12
• Local interactions make complete swarm control difficult, limiting practical applications to tasks that do not require precise dynamics and potentially increasing time and energy use.The review identifies a trade-off between robustness and efficiency.Source: C. Robotics · p. 12

D. Analytics

• Research tools for analyzing interactions and implicit communication among agents are scarce, while explicit networking may complicate adversarial reverse engineering but must be weighed against shepherding’s lack of an explicit network.Source: D. Analytics · pp. 12–13

E. Future Shepherding

• Future work could expand path and goal planning, enrich the primitive behavior sets studied for swarms and shepherds, and add more complex AI.The review says richer behaviors would increase the complexity of expressed dynamics and that more complex AI would support richer natural intelligence.Source: E. Future Shepherding · p. 13
• Future systems could combine shepherding branches in a hierarchy, with multiple UxV shepherds guiding sheep and a human monitoring and modifying behavior and trajectories remotely in real time.The human could actively modify shepherd behavior and trajectory.Source: E. Future Shepherding · p. 13
• Advances in one shepherding sub-field could transfer to another, such as adapting a robotic sheep shepherd for crowd control or vice versa.Source: E. Future Shepherding · p. 13

XI. CONCLUSION

• More than 30 years of research suggest shepherding can provide robust, flexible, scalable, simple solutions to complex tasks by assigning low-level control to the swarm and high-level planning to shepherds.The review identifies gaps spanning machine learning, artificial intelligence, robotics, and multi-agent systems, while describing these challenges as surmountable.Source: §XI · p. 13
• Most experimentation has been in simulation, although intended applications include ground and surface vehicles, robotic sheep herding, and steering bird swarms away from airports.Examples include ground vehicles navigating complex terrains and surface vehicles distributed for sea-state estimation.Source: §XI · p. 13
• Shepherding still needs study across a wider range of applications, with research gaps spanning machine learning, artificial intelligence, robotics, and multi-agent systems.The review says these challenges are not insurmountable.Source: §XI · p. 13