Dec-MARVEL lets silent drone teams explore with 16% higher success
Drones coordinate without talking, just by watching each other's paths under tight budgets.
Multi-UAV exploration often fails when communication is unreliable, sensors have limited field of view, and each robot must reserve enough battery to return home. Dec-MARVEL solves this with a completely decentralized approach: drones never exchange maps, goals, or messages. Instead, they coordinate by observing any teammate that passes within their camera's view. Each robot runs a graph-attention actor that fuses local frontier geometry, teammate motion, and remaining budget to choose waypoint-heading actions that still allow a return trip. The actor is trained with phase-conditioned critics and a mixture-based budget curriculum.
Across 900 held-out trials covering three team sizes (2, 4, 8) and three travel budgets (720, 800, 1024 meters), Dec-MARVEL matched or beat four baselines in all configurations. Under the tightest 720m budget, success rates were 53% (2 drones), 94% (4 drones), and 100% (8 drones) — compared to the strongest baseline's 37%, 83%, and 99%. It also achieved the lowest sensing overlap, meaning less redundant coverage. Physical robot experiments confirmed successful sim-to-real transfer, demonstrating real-world viability for search-and-rescue, mapping, or surveillance in communication-denied environments.
- Drones coordinate purely by visual observation of teammate trajectories — no communication needed.
- Graph-attention actor combines local terrain, teammate motion, and remaining fuel to pick return-feasible actions.
- Under 720m budget, 4-drone teams improved success from 83% to 94% over the best baseline.
Why It Matters
Enables robust multi-drone exploration in GPS-denied or communication-limited environments for search, rescue, and mapping.