This project implements a closed-loop thermal guidance and interception system inspired by real-world autonomous seekers such as infrared-guided missiles or UAV interceptors.
The objective is to autonomously detect, track, and intercept a moving heat-emitting target using only noisy sensor observations, without direct access to the target’s true state.
The system emphasizes realistic uncertainty, state estimation, and decision-making under noise, rather than hard-coded or idealized behavior.
The environment models a moving heat-emitting target whose thermal signature follows a Gaussian intensity distribution in space.
Legend
- 🔵 Agent (current position)
- 🟠 True target (ground truth)
- Dashed lines → motion history
Key characteristics:
- Continuous 2D world
- Moving target with configurable velocity
- No direct access to target state for the agent
The agent perceives the world through a limited-range thermal sensor that produces a 2D heat map centered on the agent.
Sensor limitations intentionally included:
- Finite sensing range
- Discrete spatial resolution
- Additive Gaussian noise
- Partial observability (local field of view only)
Legend
- Heat map → thermal intensity
- ❌ Cyan crosses → detections
- 🟢 Green dot → selected track estimate
This ensures the agent never “cheats” by observing global information.
From each sensor frame:
- Thermal peaks are extracted
- Noise artifacts are filtered using thresholding
- Candidate detections are generated in sensor coordinates
These detections may be:
- Noisy
- Missing
- Temporarily incorrect
This reflects realistic sensing conditions.
A constant-velocity Kalman filter is used to estimate the target’s position and velocity over time.
The tracker:
- Predicts motion during missed detections
- Corrects estimates when measurements are available
- Maintains an uncertainty covariance
Tracker uncertainty is converted into a confidence metric, which directly influences agent behavior.
The agent switches behavior dynamically based on:
- Tracker confidence
- Measurement availability
- Environmental constraints
Operational modes:
- SEARCH → low confidence, exploration
- TRACK → stable estimation
- INTERCEPT → aggressive pursuit
This mimics real autonomous decision-making systems.
When a valid target estimate exists, the agent applies Proportional Navigation (PN) guidance to compute velocity commands.
PN is widely used in real interception systems because:
- It reacts to line-of-sight rate
- It is robust to target maneuvers
- It naturally converges on moving targets
Velocity commands are bounded to reflect physical constraints.
The system runs continuously as:
Environment → Sensor → Detection → Tracking → Behavior → Guidance → Motion
Each cycle tightens the feedback loop between perception and control.
Tracker confidence is logged and plotted to analyze convergence and stability.
Observations:
- Confidence starts low during search
- Increases during stable tracking
- Peaks before interception
This explains why interception succeeds.
To ensure robustness, the system is evaluated using Monte-Carlo simulations with randomized:
- Initial target positions
- Target velocities
- Noise realizations
This confirms that interception success is not dependent on hard-coded values.
The system logs runtime events to the terminal:
Example output: 🎯 INTERCEPT ACHIEVED at step 23 Final agent position: [32.35, 25.85] Near wall: False
Logs include:
- Mode transitions
- Detection events
- Interception confirmation
This project demonstrates:
- Sensor modeling under noise
- State estimation using Kalman filtering
- Confidence-driven decision making
- Realistic guidance laws
- Modular autonomous system design
- Separation of demo and evaluation logic
It reflects core principles used in autonomous weapons, robotics, and aerospace guidance systems.