Coordinated autonomy for distributed security.
Aegis is an experimental platform for studying multi-agent aerial coordination through reproducible simulation, computer vision and multi-agent learning.
Security systems don't scale linearly.
Fixed sensors cover a fixed area. A single agent creates a single point of failure. Neither adapts when conditions change or when an agent is lost. Aegis studies whether a coordinated group of agents can cover more ground, adapt to loss, and produce results that can be measured and repeated.
Coverage is shared across agents rather than concentrated in one sensor or one point of failure.
The group redistributes itself when an agent is lost, a target appears, or conditions change.
Every scenario runs from a fixed seed and configuration, so behavior can be measured and compared.
From perception to autonomous action.
Perception is decoupled from decision-making. Each layer communicates through a fixed contract, so any layer can be replaced without changing the ones around it.
Deterministic methods first, learning second.
- YOLO
- ByteTrack
- Voronoi coverage
- Potential fields
- Multi-agent reinforcement learning
Deterministic coordination methods serve as a measurable baseline. Multi-agent learning is introduced only after a baseline exists for comparison.
Validate behavior before hardware.
Every episode is recorded in full, so any run can be inspected or repeated exactly.
Every behavior should be measurable.
Average share of the monitored area covered by the swarm over the course of an episode.
Time required to restore 90% coverage after an agent failure.
Collisions per 1,000 agent-seconds of simulation.
Packet loss applied to inter-agent communication during an episode, used as an experimental condition.
example run, not a research result
Same scenario. Different policies.
Baseline and learned policies are evaluated on identical scenarios and derived seeds, so results can be compared in pairs rather than as separate averages.
Contracts, not assumptions.
The decision layer consumes an abstract observation, not a detector's output. Perception can change without changing behavior.
Every episode runs from a fixed seed and configuration, so results can be verified and repeated.
Agents act only on what they can observe, including under partial communication.
A simulated agent and a physical one share the same interface, so validated behavior can move to hardware unchanged.
Where the project stands.
Deterministic swarm coordination and measurable coverage.
Object detection and multi-object tracking.
MARL evaluation against deterministic baselines.
Port validated behaviors to compatible aerial platforms.
Build systems that coordinate.
Aegis is an experimental platform for studying autonomous multi-agent coordination.