🎛️ Reward Function Weights
Core Mission Rewards
Formation & Coordination
Safety & Compliance
Energy & Efficiency
📊 Live Behavioral Prediction
Predicted Swarm Behavior Profile
Based on current reward weights, the trained policy will exhibit:
Aggressiveness
Medium
Risk Taking
Low
Cohesion
High
Coverage
Medium
Endurance
Medium
Safety Score
High
Reward Function Detail
Training Metrics Projection
| Metric | Predicted | Baseline |
|---|---|---|
| Convergence (episodes) | ~1.8M | 2.05M |
| Final Reward | -12.3 | -15.1 |
| Success Rate | 87% | 82% |
| Collision Rate | 0.3% | 0.8% |
| Mission Time | 4.2 min | 5.1 min |
| Battery at RTB | 28% | 22% |
⚠️ Reward Hacking Warnings
📋 Reward Function Catalog (34)
R1: Target Acquisition
r = α · 1[target_in_FOV] + β · 1[target_classified]
Positive reward when a target enters the drone's field of view and is classified by YOLOv8
R2: Area Coverage
r = γ · |new_cells_covered| / |total_cells|
Rewards exploring new grid cells on the operational map
R3: Formation Cohesion
r = -δ · Σ|pi - pcentroid|²
Penalises drones straying from the formation centroid
R4: Collision Avoidance
r = -ε · Σ 1[d_ij < d_safe]
Heavy penalty for inter-drone proximity violations
R5: Geofence Compliance
r = -ζ · 1[outside_geofence] · 100
Massive penalty for leaving the operational geofence
R6: Energy Efficiency
r = η · (1 - ΔE / ΔE_max)
Rewards energy-efficient flight paths
R7: Comm Relay
r = θ · 1[link_quality > threshold]
Rewards maintaining communication link quality
R8: ISR Dwell
r = ι · min(t_dwell, t_max) / t_max
Rewards dwelling over targets for ISR data collection
R9: Battery Reserve
r = -κ · 1[battery < 20%]
Penalises low battery states to encourage RTB
R10: Altitude Limit
r = -λ · 1[alt > alt_max or alt < alt_min]
Penalises altitude violations per airspace regs
R11-R17: Scenario-specific
Per-scenario reward shaping
SAR pattern search, CBRN plume tracking, perimeter sweep, C2 relay mesh, etc.
R18-R34: Safety + efficiency
Multi-objective shaping terms
No-fly zone, wind compensation, speed limits, path optimisation, RTB timing, etc.