Read the current state, the dated handoff
and ADR-0008 before
resuming. The no-AGPL boundary of ADR-0004 is now enforced by
tools/check_licences.py rather than remembered, and
enforcement runs in a pre-commit hook — configured, not yet activated.
Next is RF-DETR perception, starting with licence provenance: VisDrone and
UAVDT are recorded unresolved because no primary-source licence for either
exists in this repository, and the check refuses a commit that references them
until that is established. Dataset preparation and an inference baseline feeding
existing georeferencing follow.
Nothing was downloaded, trained, exported or benchmarked, no toolchain was installed and no physical-flight gate closed; legal coverage and flight approval remain PARTIAL per the drone-regulation research. Historical research figures are not new results.
Autonomous aerial security for gated residential communities.
A drone that stays on its pad until something actually happens.
-green){kind=icon}
The obvious build is a drone that flies a scheduled perimeter patrol. Every product demo in this category shows exactly that.
I costed it, and it loses.
A 5-acre society is ~20,234 m² with only ~600 m of perimeter. A drone flying 12 sorties a day at 10 minutes each is airborne 8.3% of the day — about 6.5% once you subtract monsoon and kite-season groundings. Fixed cameras bought with the same ₹5–7 lakh of capex cover 100% of it: in rain, at 03:00, silently, without overflying anyone's balcony, and without a single regulatory approval.
Per rupee of capex, a scheduled patrol drone loses to fixed CCTV roughly 12:1 on temporal coverage at the size of society that would actually buy one.
Three regulations independently punish the schedule specifically:
| Constraint | Effect on a scheduled patrol |
|---|---|
| CPCB Noise Rules 2000 — 45 dB(A) night limit in residential zones | A 2 kg quad near an occupied façade at night is non-compliant. Night patrol is out. |
| ~400 battery cycles to 80% SoH | 12 sorties/day burns ~11 battery sets a year. Patrol frequency is the opex line. |
| DPDP Act 2023 §7 — closed list, no legitimate-interest ground | Every scheduled overflight processes a resident with no lawful basis. |
So AEGIS inverts the design. The drone is not a surveillance platform — it is a response asset. Fixed AI cameras and an ESP32 perimeter-sensor grid provide the always-on layer. The aircraft launches only on a corroborated trigger, to do the things cameras cannot: reach a blind spot, follow a subject across the property, and put a steerable view on an incident for a human to adjudicate.
Full reasoning: ADR-0002.
The crossover where a patrol does win is around 20–30 acres / 1.5 km of perimeter, where trenching a camera line costs ₹25–45 lakh. That is recorded as the revisit condition, not hand-waved away.
Split strictly by latency budget. The cloud is supervise, abort, review — never fly.
┌──────────────────────────────────────────────┐
│ CLOUD (AWS ap-south-1) seconds │
│ fleet state · archive · analytics · models │
│ resident notifications · multi-site │
└───────────────▲──────────────────────────────┘
│ outbound-initiated only (CGNAT)
│ NATS JetStream leaf → hub, store-and-forward
┌───────────────┴──────────────────────────────┐
│ SITE EDGE (Ubuntu node) 100–300 ms │
│ detection · fusion · corroboration gate │
│ dispatch · takeover console · hot video │
│ ── survives total uplink loss ── │
└───────▲───────────────────▲──────────────────┘
│ WireGuard │ MQTT / LoRa
│ MAVLink 2 (signed) │
┌───────┴────────┐ ┌───────┴──────────────────┐
│ ONBOARD │ │ PERIMETER SENSOR GRID │
│ 10–50 ms │ │ ESP32 nodes · gate node │
│ flight ctrl │ │ fixed AI cameras │
│ prec. landing │ └──────────────────────────┘
│ failsafes │
│ on-frame redaction │
└────────────────┘
Why not cloud-native? Measured glass-to-glass WebRTC on India → AWS ap-south-1 → India over 4G is 740 ms P50 / 1180 ms P95. Teleoperation research puts the human-takeover ceiling at 170–300 ms. The cloud path misses by 2–4× at the median, and the dominant term is the mobile access network — no amount of backend engineering fixes it. Manual takeover is therefore an on-site capability, and the edge node keeps detecting, dispatching and recording with the uplink completely severed. (ADR-0003)
This repo documents why, not just what. Each of these was a real fork with a real cost:
| ADR | Decision | The thing that decided it |
|---|---|---|
| 0002 | Dispatch on corroborated triggers, no patrol schedule | Scheduled patrol loses to CCTV 12:1 on coverage per rupee |
| 0003 | Edge-first; cloud never in the flight loop | 740 ms measured vs a 170–300 ms requirement |
| 0004 | RF-DETR, not YOLO | Every Ultralytics generation is AGPL-3.0; "edge devices, robotics, cameras" is a named Enterprise trigger |
| 0005 | ArduPilot, run unmodified | It ships a precision-landing retry state machine; PX4 doesn't. Running stock keeps GPLv3 off our code |
| 0006 | Emit no identifiable data by default | DPDP §7 is a closed list — no legitimate-interest ground exists in Indian law |
| 0007 | ROS 2 Jazzy, bridged with MAVROS | No JetPack exists for the Ubuntu the current ROS 2 LTS needs; AP_DDS documents Humble only |
The privacy map is one signed, versioned GeoJSON artefact — society boundary plus altitude bands for every registered private aperture — that is simultaneously:
- the annexure to the RWA's approved privacy policy, and
- the runtime config that hard-slaves the gimbal and shutter.
Deployed behaviour cannot drift from approved policy, because they are the same file. Enforcement is fail-closed: on stale pose, degraded GNSS, or a signature that does not verify, the shutter closes and the gimbal stows. Degraded state means less capability, never more.
| Layer | Choice | Note |
|---|---|---|
| Flight firmware | ArduPilot Copter 4.7.x, unmodified | PLND_* retry state machine; Lua for onboard behaviour |
| Middleware | ROS 2 Jazzy + MAVROS 2.15 | Jazzy is the only LTS matching JetPack 7 (Ubuntu 24.04) |
| Companion | Jetson Orin Nano Super 8 GB | 67 TOPS; runs detection and a 30 Hz landing loop concurrently |
| Detector | RF-DETR (Apache-2.0) | RF-DETR-S: 53.0 COCO AP @ 512 px, 3.5 ms T4-FP16 |
| Small objects | SAHI sliced inference (MIT) | +5–7 AP at 4–6× compute — sweep mode, not every frame |
| Tracking | Roboflow trackers + supervision |
Apache-2.0 / MIT; not BoxMOT (AGPL) |
| Simulation | Gazebo Harmonic + ArduPilot SITL | AirSim is dead; Colosseum archived 2026-07-11 |
| Edge | Docker Compose, MediaMTX, NATS JetStream leaf | Store-and-forward through partitions |
| Cloud | FastAPI, Postgres 17 + PostGIS, NATS hub | Partitioned Postgres beats a TSDB below ~10k pts/sec |
| Web | React 19 + MapLibre GL + deck.gl | |
| Sensor nodes | ESP32-S3 | Perimeter, gate, dock controller |
Rejected alternatives and the reasons are in the ADRs — including the licence
traps that a pip install does not surface (RF-DETR XL is PML 1.0, not
Apache-2.0; D-FINE's Objects365 checkpoints are not commercially cleared even
though the repo is Apache-2.0).
| Document | What it is for |
|---|---|
docs/STATE.md |
Start here. Where the project is, what changed, what to do next |
docs/adr/ |
Why each decision was made, and what would make us revisit it |
docs/research/ |
The sourced research the decisions rest on — every claim carries a URL and a confidence level |
docs/ops/setup.md |
Environment setup; says which toolchains you actually need |
CONTRIBUTING.md |
Conventions, and the six rules that are not negotiable |
CHANGELOG.md |
What has been decided, added and fixed |
| Note | Covers |
|---|---|
| architecture | Edge/cloud split, transport, video, alerting, Indian carrier constraints |
| cv-models | Detector licensing, small-object detection, edge accelerators, georeferencing |
| flight-stack | ArduPilot vs PX4, ROS 2, precision landing, docking, simulation |
| privacy-law | DPDP Act, consent, overflight, CERT-In, case law |
| prior-art | Commercial drone-in-a-box landscape and the India gap |
| unit-economics | What societies pay today, cost to serve, noise, insurance, risk register |
| hardware-bom | Priced BOM at three tiers, and the RF licensing trap |
Early and honest. This section tracks reality, not intent.
- Architecture decided and documented — 7 ADRs
- Domain model: typed ULID identifiers, datum-safe geodesy
- Privacy map: schema, HMAC signing, 30 Hz gimbal constraint solver
- Corroboration gate: modality-family fusion, admission gates, audit trail
- Georeferencing: camera model, damped-Newton undistortion, error budget
- Perception: dataset pipeline, RF-DETR fine-tune, ONNX/TensorRT export
- Autonomy: ROS 2 nodes, sortie state machine, SITL harness
- Simulation: residential-society world
- Backend services and ops dashboard
- ESP32 sensor-node firmware
- Hardware design spec and BOM
- Paper and deck
Scope note. Validation is in high-fidelity simulation — real ArduPilot firmware, real ROS 2, real models trained on real aerial datasets, benchmarked on real hardware. The airframe and dock are specified and costed as an engineering design; no physical aircraft has been fabricated. Every number quoted in this repository is reproducible from this repository.
make doctor # check which toolchains you have, and how to get the rest
make setup # bootstrap Python and web dependencies
make test # run the suites
make help # everything elseThe repo spans four toolchains that share no package manager (colcon, pip, pnpm,
PlatformIO), so make is the single entry point regardless of which corner of
the tree you are standing in.
aegis/ shared library — domain model, privacy map, dispatch, vision
domain/ ids, geodesy, privacy map, sensing, corroboration gate
vision/ camera model, georeferencing
autonomy/ ROS 2 workspace: flight, sortie state machine, SITL harness
perception/ dataset pipeline, training, evaluation, export
services/ edge and cloud backend services
web/ operations dashboard
firmware/ ESP32 perimeter, gate and dock-controller firmware
sim/ Gazebo worlds, ArduPilot SITL configuration
infra/ compose, k8s, observability
tests/ 183 tests, property-based where the maths warrants it
docs/
STATE.md where the project is, and what to do next — read this first
adr/ architecture decision records — the reasoning
research/ sourced research notes behind every decision
ops/setup.md environment setup for all four toolchains
paper/ IEEE-format paper
Apache-2.0. See LICENSE.
No AGPL-licensed code enters the inference or training path, enforced by tools/check_licences.py rather than by memory: forbidden packages are a denylist, and every model weight or dataset reference must resolve to a reviewed registry entry or it is refused. It runs from make lint and from a pre-commit hook that is configured but not yet activated — see the current checkpoint above. (ADR-0004, ADR-0008)