NETRA

Patrol command center for Pashupati robots: the live fleet on the mall map, patrols and schedules, AI anomaly detection with operator validation and Telegram alerts, crowd analytics, reports, and a read-only API for third-party applications. By EZRA Robotics.

Features · System · Run on a Mac · Deploy · Integration API · Documentation · Develop

---

Robots patrol mall corridors on recorded routes and inspect the floor at their stop points. NETRA is where people watch them, start and schedule patrols, and decide what the robots found. This repository is the deployment root: the infrastructure at the top level, the application in two submodules.

Features

Monitor

• Dashboard for the control room wall: today's figures, a live map with robot trails and follow mode, one queue of what needs attention now, the fleet with its controls, the activity log.
• Calibrated mall floor plan (SVG, PNG, JPG or PDF), or a metric grid in the robot's map frame until there is one. Robots, routes, anomalies and the crowd show on either.
• Robot health at a glance: battery, protective field, localization, CPU, GPU, RAM and temperature, with history charts and a replay of the track.

Patrol

• Start a patrol now, or on weekly and one-off schedules, on any route the robots recorded.
• Pause, resume, stop, localize, and a latched e-stop that only an explicit release lifts.
• Turn the robot's AI inspection and people detection on or off, and choose the model and its pace, from the browser.

Anomalies

• The robot's early detection is checked again by a larger vision-language model on the server.
• Per category: confirm at once and alert a Telegram group (trash to the cleaning crew), or wait for an operator to confirm or reject.

Analyse

• Crowd heatmaps and hourly and daily peaks from the people the robots see.
• PDF and Excel reports per patrol and per period, in English or Indonesian.

Integrate

• Third-party applications read anomalies (with their original photos), robots, patrols, routes and floor plans over OAuth2 client credentials, with an OpenAPI description and Swagger UI.

Operate

• Roles (viewer, operator, supervisor, admin), an audit log of every change, login throttling.
• Runs on a VPS with HTTPS, or on a Mac on the robots' network for a demo or a site without internet; Prometheus and Grafana included, daily database backups.

System

flowchart LR
    subgraph Site[Mall]
        R[Genisom L1W<br/>Pashupati stack] <-- Wi-Fi --> H[Handheld<br/>Pashupati Controller]
    end
    R <-- WebSocket + HTTPS --> C[Caddy<br/>HTTPS]
    B[Browser<br/>operators] --> C
    P[Third-party app] -- OAuth2 --> C
    C --> API[netra-api<br/>Go]
    C --> WEB[netra-web<br/>Next.js]
    API --> DB[(TimescaleDB)]
    API --> NATS[NATS JetStream]
    API --> AI[netra-ai<br/>FastAPI] --> OL[Ollama<br/>vision-language model]
    API --> TG[Telegram]
    PR[Prometheus + Grafana] -.-> API

Loading

Path
be/	netra-api (Go) and netra-ai (FastAPI, vision-language model)
fe/	netra-web (Next.js dashboard)
docker-compose.yml	the whole stack
Caddyfile	HTTPS, /api to the API, /grafana to Grafana, the rest to the dashboard
monitoring/	Prometheus scrape config, Grafana datasources (Prometheus, TimescaleDB) and dashboard
scripts/deploy.sh	server: pull, update submodules, pull images, restart, pull the AI model
docker-compose.local.yml, Caddyfile.local, .env.local.example, scripts/local.sh	running on a Mac (LOCAL.md)
docker-compose.tunnel.yml, Caddyfile.tunnel	a Mac behind a Cloudflare Tunnel, LAN port kept for robots
docs/	operator guide (PDF) and its LaTeX source

The robots run Pashupati; the handheld is Pashupati Controller.

Run on a Mac

Robots on the same network as the Mac. See LOCAL.md: scripts/local.sh setup && scripts/local.sh up, then scripts/local.sh robot <id> prints the robot's connection settings. For a demo on a laptop, scripts/local.sh export on the server and scripts/local.sh import on the laptop copy the data over, so robots keep their tokens and only change CC_URL (Demo: the command center on a laptop).

Deploy on a server

git clone --recursive git@github.com:DarmawanWeb/Netra.git && cd Netra
cp .env.example .env            # domain, passwords, Telegram, AI model
docker compose up -d --build    # or with CI images: docker compose pull && docker compose up -d
docker compose exec ollama ollama pull qwen2.5vl:7b

Then open https://<domain>, log in as NETRA_ADMIN_USER, add users (roles viewer, operator, supervisor, admin), register robots and upload the floor plan. Grafana is at /grafana.

Updates, backups

Updates: scripts/deploy.sh (images from GHCR) or scripts/deploy.sh --build. The deploy workflow runs it over SSH on v* tags (secrets DEPLOY_HOST, DEPLOY_USER, DEPLOY_KEY, DEPLOY_PATH).

Backups: daily pg_dump into backups/ (kept BACKUP_KEEP_DAYS); restore with docker compose exec -T db pg_restore -U netra -d netra --clean < backups/netra-YYYY-MM-DD.dump. Anomaly photos and floor plans are in the api-data volume.

Integration API

Other systems read the anomalies (with their original photos), robots, patrols, routes and floor plans over https://<domain>/api/ext/v1, with OAuth2 client credentials. An admin creates a client in Settings → API clients (scopes; the secret is shown once) and gives the third party the client_id, the client_secret, the guide INTEGRATION.md and the interactive docs at https://<domain>/api/ext/docs (OpenAPI: /api/ext/v1/openapi.yaml).

Documentation

For	Document
Operators, security staff, supervisors	Panduan Operator (PDF, Indonesian): safety, daily start with the Genisom L1W, dashboard, patrols, validation, schedules, routes, reports, troubleshooting
Running NETRA on a Mac	LOCAL.md
Backend developers	netra-fleet-be/docs: architecture, dashboard API, database, configuration, development, AI service
Dashboard developers	netra-fleet-fe/docs: architecture, development
Third-party integrators	INTEGRATION.md and /api/ext/docs on the server
Robot side	Pashupati (robot API)

Develop

git submodule update --init --recursive
lefthook install          # in the root and in each submodule

The validate workflow checks the compose files, the Caddyfiles and the shell scripts on every push; each submodule has its own CI (lint, tests, build) that publishes its images to ghcr.io/darmawanweb on main and v* tags. See be/README.md and fe/README.md.

Released under the Apache License 2.0.