Simulation prototype
Omnivision
A drone fleet that keeps working when the link doesn’t.
Peer-to-peer coordination, edge AI and legacy adapters in one stack, for search, survey and inspection missions.
Design targets for the simulation prototype, not field results. Scroll to explore.
Seen from above
Built for the day the link drops
Observation and reporting only. No weapons or effector interfaces. An operator can pause, override or recall at any time.
The problem
Capable but fragile, or robust but limited
Today's teams are forced to choose. Omnivision is a software layer that removes the trade-off: one coordination fabric, the best AI the hardware can afford, and graceful degradation instead of failure.
The link is a single point of failure
Most uncrewed systems lean on a ground station. When the link degrades or a vehicle drops out, missions fail or vehicles simply return home.
Modern AI is too heavy for the airframe
State-of-the-art detection models outgrow the small onboard computers most drones actually carry.
Legacy protocols make change painful
Existing platforms speak older or proprietary protocols, so adding new software means replacing hardware.
The demo
One mission. Everything goes wrong.
A fleet of 4–6 simulated UAVs searches a defined area for people, vehicles and hazard markers, while we break it on purpose.
- 1
Normal operations
The fleet divides the area, each vehicle flies its search pattern and detections are shared. The operator sees a live map and detection feed.
- 2
Comms degradation
Latency spikes and packet loss rise. The fleet keeps coordinating at reduced update rates.
- 3
Network partition
The fleet splits into two groups. Each elects its own captain and carries on, then merges world models on reconnection.
- 4
Captain loss
The current captain is taken out. A new one is elected within the target time and the mission continues.
- 5
Component failure
A vehicle loses its camera or main compute and drops to a degraded mode: a smaller model, or relay-and-support duty.
- 6
Mission complete
Coverage finishes and the fleet produces a report: coverage %, detections and a timeline of every failure and recovery.
Architecture
Five layers, one stack
Each area is a separate package with a documented interface. Any vehicle can serve as the gateway to the operator.
Operator dashboard + mission report
Live map, vehicle health, captain identity, comms graph, detections. Pause, override or recall at any time.
React · map library
Fleet fabric
Discovery, captain election, decentralized task allocation, world-model merge and health monitoring.
Zenoh / DDS · Raft-style election · CBBA / auctions · CRDTs
Per-vehicle autonomy agent
Local mission plan, degradation mode manager and the edge AI cascade.
ROS 2 · ONNX Runtime · INT8 / pruning / distillation
Vehicle abstraction + adapters
One small, typed vehicle interface in front of every platform.
MAVLink (pymavlink / MAVSDK) · mock legacy protocol
Simulation + fault injection
Multiple simulated UAVs with cameras, and on-demand faults: latency, loss, partitions, node death.
PX4 SITL · Gazebo · tc/netem · Docker
Autonomy under loss
Degrade on purpose, not by accident
Every vehicle carries a local mission plan and an explicit policy for loss of comms, peers or components. Faults trigger clean, testable transitions, never improvisation.
Continue mission
Comms lost, but the local plan is intact. Keep searching and buffer detections.
Hold and relay
Hover at a useful point and bridge messages between separated peers.
Reduced-capability search
Camera or compute degraded. Switch to a smaller model and a narrower task.
Return to rally point
When nothing useful remains, recover safely to a pre-agreed location.
Success metrics
Numbers we hold ourselves to
Starting targets for the simulation prototype, calibrated after baseline runs. Design targets, not field results.
Mission completion
Nominal runs reaching required coverage, across 30+ seeded runs per scenario. ≥ 75% under injected faults.
Captain re-election
From captain loss to a new captain accepted by all reachable peers.
Recovery after vehicle loss
Until a lost vehicle's tasks are reassigned to the rest of the fleet.
Detections lost after merge
World models converge after a partition heals, with nothing dropped.
Faster edge inference
Per-frame latency versus the baseline model, under emulated edge limits.
Accuracy retention
mAP / F1 of the compressed model versus the uncompressed baseline.
Frames on the light model
Share of frames the cascade handles without waking the heavy model.
Adapter fidelity
Command and telemetry round-trip success on the defined test suite.
Where it matters
Built for places the network forgets
Disaster response
Collapsed structures, flooded valleys, wildfire edges, where infrastructure is the first thing to go.
Search and rescue
Wide-area search for people that can't stop because one radio dropped.
Infrastructure inspection
Pipelines, power lines and industrial yards with patchy coverage.
Perimeter monitoring
Persistent observation and reporting over large, remote sites.
Roadmap
From simulation to the sky
A pilot track with disaster-response, inspection and infrastructure partners runs in parallel, on the same core architecture.
NOW
Simulation prototype
4–6 simulated UAVs, fault injection, edge benchmarks and a working MAVLink adapter.
PHASE 2
Hardware-in-the-loop
A low-cost flight controller and companion computer, real edge benchmarks, a second real adapter.
PHASE 3
Small physical fleet
A controlled, legally approved test environment, real radio mesh and GPS-denied navigation experiments.
PHASE 4
Hardening
Security hardening, third-party review and certification paths with an experienced partner.
Principles
What we will and won't build
Written down before we pitch anyone. It protects the team and makes us a more credible partner.
Observation and reporting only
No weapons, payload-release, target-selection or engagement logic. No interface to any effector.
A human can always step in
Whenever a link exists, the operator can pause, override or recall any vehicle or the whole fleet.
No offensive capability
Network degradation is used only to test our own resilience. No jamming, no offensive cyber.
Honest numbers
Seeded, reproducible runs, published methods and failure cases. A research architecture, not "military grade".
Work with us
Help us keep the fleet flying
We're looking for mentors, pilot partners and funders in disaster response, inspection and infrastructure. See the demo, break it, and tell us what you need to say yes.