9 min

Autonomous Navigation & Swarms

Sensor Fusion, GPS-Denied Flight & Multi-Drone Coordination

Why GPS Alone Fails

GPS provides 2-5 meter horizontal accuracy under ideal conditions. In real operations, those conditions rarely hold. Urban canyons in Delhi or Mumbai create multipath reflections where satellite signals bounce off buildings, producing position errors of 10-50 meters. Dense forest canopy in the Western Ghats attenuates signals to unusable levels. Electromagnetic interference near power lines and industrial zones corrupts receiver lock. Deliberate GPS spoofing — broadcasting fake signals to hijack the drone's position estimate — is a demonstrated threat that militaries and hostile actors can deploy.

Any drone system that depends solely on GPS for navigation will fail in precisely the scenarios where autonomous operation matters most: contested environments, infrastructure-poor regions, and dense urban areas. Sensor fusion is the answer.

Sensor Fusion Architecture

The goal: combine noisy, complementary sensor measurements into a single state estimate (position, velocity, orientation) that is more accurate and robust than any individual sensor.

Sensor	Measures	Update Rate	Strengths	Weaknesses
GPS	Lat/Lon/Alt absolute position	5-20 Hz	Global, drift-free	Multipath, denial, 2-5m accuracy
IMU (Accelerometer + Gyro)	Linear acceleration, angular rate	200-1000 Hz	Very fast, works everywhere	Drift accumulates (gyro bias, accel noise)
Barometric Altimeter	Pressure altitude	10-50 Hz	Independent altitude source	Weather-dependent drift, relative not absolute
Optical Flow	Ground-relative velocity	30-60 Hz	Works without GPS, detects drift	Fails over featureless terrain (water, sand)
LiDAR	Range to obstacles/ground	10-100 Hz	Precise distance, works in dark	Heavy, expensive, limited range
Visual-Inertial Odometry (VIO)	Relative pose from camera + IMU	30 Hz	Lightweight, rich environmental info	Drift over long distances, lighting-dependent
Magnetometer	Heading (magnetic north)	50-100 Hz	Absolute heading reference	Susceptible to magnetic interference

Extended Kalman Filter (EKF)

The workhorse of sensor fusion. The EKF maintains a probabilistic state estimate — not just "I am at position X" but "I am at position X with uncertainty covariance P." At each timestep:

Predict — use the IMU measurements and a motion model to propagate the state forward (fast, every IMU sample)

Update — when a GPS fix, barometric reading, or optical flow measurement arrives, correct the prediction using the Kalman gain (slower, at sensor rate)

The Kalman gain automatically weights each measurement by its reliability. If GPS uncertainty spikes (poor satellite geometry), the filter relies more on IMU propagation. If IMU bias has drifted, a clean GPS fix pulls the estimate back.

Complementary Filters

Simpler than EKF but effective for specific sensor pairs. A classic example: fuse gyroscope (accurate short-term, drifts long-term) with accelerometer-derived attitude (noisy short-term, correct long-term) using a high-pass/low-pass frequency split. Computationally cheap enough for microcontrollers on small drones.

GPS-Denied Navigation

When GPS is completely unavailable, the drone must navigate using:

Visual-Inertial Odometry (VIO) — track visual features frame-to-frame and combine with IMU to estimate motion. State-of-the-art VIO (ORB-SLAM3, VINS-Fusion) achieves <1% distance error over short trajectories but drifts over kilometers.

Terrain-Relative Navigation (TRN) — match the terrain profile seen by downward-looking LiDAR or radar against a stored Digital Elevation Model. Provides absolute position fix without GPS. Works best over distinctive terrain; fails over flat featureless plains.

Visual Place Recognition — match current camera view against a pre-built visual database of the operating area. AI-based descriptors (NetVLAD, SuperGlue) handle viewpoint and lighting changes that break classical feature matching.

Celestial Navigation — for high-altitude long-endurance platforms, star trackers provide absolute attitude and position. Not practical for low-altitude multirotors.

Prompt: "A drone operating in GPS-denied conditions over urban Bengaluru has VIO, downward LiDAR,
barometric altitude, and a pre-loaded 3D building map. After 2 km of flight, VIO drift is
estimated at 15 meters. Describe a sensor fusion strategy to bound position error to <5 meters
using building-edge matching against the map as a correction source."

Open data/sensor-fusion-log.csv — it contains synchronized sensor streams (GPS, IMU, barometer, optical flow) from real drone flights, including segments with intentional GPS denial. Use this to experiment with fusion algorithm performance under degraded conditions.

Swarm Coordination

A drone swarm is more than multiple drones flying simultaneously. It is a distributed system where coordination creates capabilities no single drone possesses.

Task Allocation

Given N drones and M tasks (survey zones, delivery points, search sectors), the allocation problem minimizes total mission time or energy while respecting constraints:

Allocation Method	How It Works	Best For
Centralized auction	Ground station collects bids from all drones, assigns optimally	Small swarms (<10), reliable comms
Distributed market	Drones negotiate peer-to-peer, trade tasks for efficiency	Medium swarms, latency-tolerant
Consensus-based (CBBA)	Each drone builds a task bundle, resolves conflicts via consensus	Large swarms, limited bandwidth
Reinforcement learning	Learned policy maps state to task selection	Dynamic environments, complex objectives

Collision Avoidance

With 5+ drones in close proximity, collision avoidance is critical:

Velocity Obstacles (VO) — each drone computes the set of velocities that would lead to collision with any neighbor and chooses a velocity outside this set

ORCA (Optimal Reciprocal Collision Avoidance) — extension of VO where each drone assumes the other will share the avoidance burden equally. Produces smooth, efficient trajectories

Priority-based deconfliction — assign priority levels (e.g., lower battery = higher priority). Lower-priority drones yield to higher-priority ones. Simple and predictable

Communication Topology

Topology	Structure	Bandwidth Need	Failure Mode
Star	All drones talk to central base	High at base	Base failure = total loss
Mesh	Each drone talks to neighbors	Distributed	Graceful degradation
Hybrid	Mesh between drones, periodic base sync	Moderate	Resilient

Failure Handling

When one drone in a swarm fails (motor failure, communication loss, forced landing):

Detect — heartbeat timeout (typically 3-5 seconds of no communication)

Re-allocate — redistribute the failed drone's remaining tasks to surviving members

Re-plan — adjust formation or coverage pattern to maintain mission objectives

Safe the failed unit — if communication is still possible, command return-to-home or controlled descent

Indian Applications

Agricultural Spraying Swarms

Large-scale spraying operations in Maharashtra and Punjab use 5-drone formations covering 50 acres per battery cycle:

Formation geometry — parallel line abreast with 5-8 meter spacing (matched to spray swath width)

Wind-adaptive spacing — AI adjusts inter-drone distance based on real-time wind to prevent spray overlap gaps

Endurance management — stagger battery changes so only one drone is grounded at a time, maintaining 80% coverage continuity

Obstacle handling — if one drone encounters a tree line or power line, it communicates the obstacle to trailing drones for pre-emptive avoidance

Regulatory compliance — DGCA requires visual line of sight for each drone unless BVR-II certified. Swarm operations currently need one pilot per drone or special exemption

Flood Survey in Bihar

The Kosi and Gandak river basins in Bihar experience devastating annual floods. Swarm survey capabilities include:

Rapid area coverage — 5 drones can survey 20 sq km in 2 hours, versus 2-3 days by foot team

Real-time mapping — onboard processing generates orthomosaic maps during flight, transmitted to NDRF operations centres

Victim detection — AI-based human detection in thermal imagery, prioritizing rescue dispatch

Communication relay — drones at altitude serve as mesh network nodes when ground cellular infrastructure is destroyed

Border Surveillance Concepts

India's 15,000+ km of land borders include terrain ranging from Rajasthan desert to Himalayan passes to Sundarbans marshland. Swarm surveillance concepts under development involve:

Persistent coverage — relay-style operations where fresh drones replace low-battery units seamlessly

Anomaly detection — AI identifies movement patterns inconsistent with normal civilian activity

Terrain-adaptive formation — swarm automatically adjusts altitude and spacing based on terrain type and threat level

Open data/waypoint-data.json for mission waypoint sets used in multi-drone operations, including individual drone assignments within formation flights.

Open data/swarm-coordination-scenarios.json for scenario definitions — task lists, drone capabilities, communication constraints, and failure injection events — designed for testing swarm coordination algorithms.

Key Takeaways

Sensor fusion is non-negotiable for operational drones — GPS alone will fail in the environments that matter most. An EKF or similar filter combining IMU, barometer, optical flow, and GPS creates robust navigation that degrades gracefully.

GPS-denied navigation is solved for short ranges — VIO + terrain-relative matching provides sub-5m accuracy over typical mission distances. Long-range GPS-free navigation remains a research challenge.

Swarm coordination is a distributed systems problem — task allocation, collision avoidance, and failure handling must work with limited bandwidth and without central authority for true operational resilience.

Indian use cases demand swarm capability — agricultural scale, flood survey urgency, and border surveillance persistence all exceed single-drone capacity. The regulatory framework (DGCA swarm exemptions) is the pacing constraint, not the technology.

This is chapter 5 of AI for Aerospace & Drones.

Get the full hands-on course — free during early access. Build the complete system. Your projects become your portfolio.

View course details

Ch. 4: Aerodynamic Simulation with AI

Ch. 6: Space Mission AI Applications