GPS-Denied · EW-Contested · Sovereign

VEGA AUTONOMY STACK by Rayfik Robotics

One lens. Complete independence. Millimetric terminal guidance, visual localization, and obstacle-aware navigation — driven entirely by monocular RGB in GPS-denied, EW-contested airspace.

Explore the Stack Mission Profiles
ACQUIRING...
VEGA · PERCEPTION ACTIVE
MODE: VISUAL_SLAM
DEPTH: MONOCULAR/METRIC
GPS: DENIED · EW: CONTESTED
CONF: 97.4%
LAT: 28.6139° N
FRAME: 00441
VEGA
v1.0
< 30ms
End-to-End Latency
< 0.5m
Terminal CEP
< 5m
GPS-Denied Localisation
0
ITAR / Foreign Dependencies
01 // Perception Primitives

THE EYES
OF VEGA

Everything begins with a single uncooled RGB sensor. No LiDAR. No radar. No active emission. Pure passive intelligence, processed at the edge.

01 — DEPTH
Metric-Scale Monocular Depth
Self-supervised depth networks fine-tuned on sovereign Indian terrain data. IMU-anchored scale recovery eliminates the classic monocular ambiguity.
RMSE < 0.5m @ 50m AGL · 30 FPS on Jetson Orin
02 — ODOMETRY
Visual Odometry
SuperPoint feature tracking fused with dense optical flow. IMU preintegration via factor graphs. Works in featureless Thar desert and high-texture Himalayan terrain alike.
Drift < 1% over 1 km · Feature-sparse capable
03 — SEMANTICS
Semantic Scene Understanding
Lightweight SegFormer with MobileViT backbone. Classifies terrain type, threat objects, and landing zones in real-time — fully on-board, no datalink.
15 FPS on-edge · 12 terrain classes · Adversarially robust
04 — STABILITY
Electronic Image Stabilisation
IMU gyro fused with optical flow compensation. Targets moving at 150+ km/h remain sharp across high-vibration CTOL airframe configurations.
Sub-pixel registration · 150+ km/h target stability
05 — PASSIVE
Zero RF Emission
VEGA emits nothing. No laser rangefinder. No active radar. Passive electro-optical acquisition makes the platform electronically silent during the entire engagement.
RF-silent · SIGINT-transparent · Low observable
06 — DATA
Sovereign Dataset Pipeline
All training data collected over Indian terrain — Thar, Himalayan treeline, Deccan plateau, urban theatre. ITAR-free ground truth via domestic survey aircraft.
10k+ annotated frames · Multi-terrain · ITAR-free
02 // Architecture

HARDWARE-AGNOSTIC
BY DESIGN

A clean HAL separates algorithms from silicon. The same perception graph runs on Jetson Orin, Snapdragon Flight Pro, or any future edge SoC — algorithm unchanged.

Mission
Intercept Behavior ISR Loiter Contingency Mgmt Abort Envelopes AES-256 GNC
↓ guidance commands ↓
Guidance
IBVS Terminal Homing Proportional Navigation MPPI Replanning iLQR Optimisation Kinodynamic Constraints
↓ state estimates ↓
Localisation
DROID-SLAM Terrain-Relative Nav iSAM2 Factor Graph Loop Closure EW Blackout Recovery NetVLAD Re-localisation
↓ features + depth ↓
Perception
Monocular Depth Visual Odometry YOLOv9 Detection ByteTrack Semantic Segmentation EIS Stabilisation
↓ raw pixels + imu ↓
Sensor / HAL
Monocular RGB (any) IMU / MEMS Barometer Magnetometer Hardware Abstraction Layer ROS 2 / DDS
03 // Mission Profiles

ONE STACK.
TWO MISSIONS.

Shared perception and localisation core. The mission layer is the only thing that changes. 80% identical codebase — two completely different terminal behaviours.

Profile Alpha
Precision Interceptor
Autonomous terminal guidance for loitering munitions and kinetic interceptors. IBVS-driven control surfaces achieve sub-meter CEP without any active RF emission whatsoever.
  • IBVS homing on tracked target centroid
  • YOLOv9 + Siamese network under partial occlusion
  • PnP geometry — sub-0.3° bearing error at 200m
  • Proportional Navigation with maneuver prediction
  • Counter-camouflage via spectral and texture analysis
  • Target lock acquired in under 1 second
CEP < 0.5m @ 300m slant range
Profile Bravo
ISR / Persistent Surveillance
Wide-area surveillance with geo-referenced target cueing, change detection, and autonomous loiter — entirely GPS-denied, no datalink required.
  • Persistent multi-target tracking with re-ID
  • Geo-referenced coordinates via TRN + Visual SLAM
  • Change detection between successive passes
  • Obstacle-aware loiter in denied airspace
  • On-board classification — no cloud dependency
  • Operates through GPS spoofing and RF jamming
Geo-location < 5m CEP GPS-denied
04 // Sovereignty

NO FOREIGN
KILL-SWITCH.

Every weight, every dataset, every algorithm — developed and owned domestically. VEGA operates under complete Indian sovereign control.

0
ITAR / EAR Components
100%
Domestically Trained Models
0
Active RF Emissions
∞
Operational Independence
05 // Development Roadmap

36 MONTHS
TO FLIGHT

01
M0 – M12
Perception Primitives
Monocular depth, visual odometry, and the sovereign dataset pipeline. The bedrock of the entire stack.
Metric DepthVisual Odometry IMU PreintegrationTerrain DatasetsHAL Design
VO drift < 1% · Depth RMSE < 0.5m
02
M6 – M18
GPS-Denied Localisation
Visual SLAM, terrain-relative navigation, and recovery from EW blackout. Localise anywhere on the subcontinent without GPS.
DROID-SLAMTerrain-Relative Nav iSAM2Loop ClosureEW Recovery
CEP < 5m GPS-denied over 10 km
03
M12 – M24
Obstacle-Aware Navigation
Time-to-collision detection, MPPI replanning, and real-time trajectory optimisation for high-speed CTOL airframes.
TTC DetectionMPPI iLQRSemantic SegContingency Mgmt
Replan < 80ms · Avoidance @ 120 km/h
04
M18 – M30
Terminal Guidance
IBVS-driven precision homing. Detection, tracking, PnP geometry, and proportional navigation converge into millimetric accuracy.
YOLOv9Siamese Tracker IBVSPnP GeometryProNav
CEP < 0.5m @ 300m · Lock < 1s
05
M24 – M36
Integration & Certification
HITL/HWIL validation, EW hardening, adversarial robustness training, and CEMILAC airworthiness. First live autonomy flight.
HITL SuiteEW Hardening DO-178C Level BCEMILACTri-Service Demo
First live flight · CEMILAC SoA issued
06 // Engage

READY FOR
INTEGRATION?

Rayfik Robotics briefs qualified defense primes, DRDO labs, and IDEX-funded ventures. Technical briefing available under NDA.

Request SDK Access ↗