When Your Satellite Window Is 4 Minutes
The Navy's unmanned vessel program had a problem: their ML models needed updates, but the vessels were underwater. Satellite links cost $14/minute and dropped constantly. Cloud-based ML wasn't going to work.
See How It Works- Client
- U.S. Navy
- Industry
- Defense & Public Sector
- Use Case
- Edge ML Analytics for Unmanned Maritime Vessels
- Timeline
- Initial vessel in 11 weeks, fleet rollout ongoing
- ROI
- Eliminated dependency on real-time connectivity
The Challenge
The unmanned vessel program was stuck. ML models trained on shore worked great - until you tried to update them on vessels operating in the Western Pacific. The existing approach required vessels to surface, establish satellite link, and maintain connection for the full update. Success rate was around 23%.
- 01 Satellite bandwidth runs $14/minute and connections drop mid-transfer
- 02 Vessels surface for 4-6 minutes on average before diving again
- 03 Failed updates meant vessels ran stale models for weeks
- 04 DoD security review for any new software takes 9 months minimum
- 05 Existing ML platform required constant cloud connectivity
- 06 Program was 14 months behind schedule on autonomy milestones
The Solution
We built a system that assumes the connection will fail. Model updates break into small chunks. Vessels grab what they can in each window. The orchestrator tracks what each vessel has and what it still needs. A full model update completes across 3-4 surface windows instead of requiring one long session.
Chunked Model Delivery
Model updates split into 200KB segments. Each chunk verifies independently. Vessels resume from last successful chunk - no wasted bandwidth on retransmission.
Fleet-Wide State Tracking
Shore command sees exactly which models each vessel has, when they last connected, and what updates are queued. Priority vessels get updated first.
Opportunistic Data Return
Vessels collect sensor data continuously. When they surface, high-priority data uploads first. Raw feeds compress and transfer during longer windows. Nothing gets lost.
The Results
The program caught up on its autonomy milestones. Model update success rate went from 23% to 97%. The 9-month DoD security review came back clean - they appreciated that we assumed hostile networks from the start.
- Model update success rate jumped from 23% to 97%
- Full fleet receives updates within 72 hours of release
- First operational vessel deployed in 11 weeks
- DoD security review passed without findings
- Satellite costs dropped 34% - less retransmission, smaller payloads
- Vessels now run 3 concurrent ML models instead of 1
- Program recovered 14-month schedule slip in 6 months
Deploying to disconnected environments?
If your edge devices can't maintain constant connectivity, we should talk. We've deployed on vessels, aircraft, and remote sites where the network is hostile by design.