Robots are getting cheap; operating a mixed fleet of them is not. Yantrika is an open-source, out-of-the-box operations layer for AMR fleets: a VDA 5050 wire in, a Supabase/PostgREST backend in the middle, and on top a live web console, automatic incidents, predictive maintenance, real alert delivery, and an AI copilot that only answers from your fleet's actual data. Every robot-affecting action passes a human approval gate. It runs end-to-end on your laptop in about 60 seconds — no cloud account, no API keys, no robots required.
python -m yantraops up --loopback
The whole platform — simulated fleet, detector, notifier, copilot, console — against an embedded in-memory backend. One process tree, one Ctrl-C.
Modules
Every module reads and writes the same Supabase tables and the same canonical status vocabulary. Each is independently installable and offline-tested.
The canonical status contract: one vocabulary (active · idle · charging ·
paused · estop · degraded · fault) plus the multi-site id helper, enforced
end-to-end by a DB CHECK constraint. Small on purpose.
Deterministic 10-AMR warehouse simulator (3 vendors, waypoint grid, chargers, faults, rolling missions). Emits real VDA 5050 v2.1 state to Supabase, MQTT, or stdout — the stand-in until your robots show up.
The VDA 5050 ⇄ backend connector: MQTT or JSONL state in, deduped
rows/alerts out; imports existing MCAP/JSONL recordings; publishes approved
operator commands back to robots as VDA instantActions.
Auto-incidents with PagerDuty-style dedup, pending windows, flap counting and
idempotent INC-XXXX ids — plus a predictive-maintenance mode that
turns telemetry trends into findings with rough RUL estimates.
Alert and incident fan-out to Slack, Discord and JSON webhooks: retries with
backoff, a no-loss pending queue, per-channel circuit breakers, severity-grouped
digests, and a 10-second test subcommand.
Grounded AI copilot with three degradation tiers: LLM + live fleet tools → LLM only → offline template answers (no key needed). A grounding guard checks the numbers in every answer against tool payloads. Also an MCP server.
The one-command orchestrator: up --loopback (zero-config demo),
--supabase, --mqtt; migrate applies the
schema with tracking and checksums; doctor preflights your
environment; status pings a running stack.
A single HTML file, no build step: live map, robot drill-down, the command approval gate, incident replay from recorded telemetry, printable shift report, command palette, and a first-run tour. Falls back to an in-browser sim offline.
Operator training as a static single-file app, served next to the console at
/academy/ by yantraops up and the AWS deploy.
Open the Academy.
Architecture
Robots (or the simulator) publish VDA 5050 state over MQTT; the bridge turns it into rows; everything else reads those rows over PostgREST. Nothing touches a robot without an operator-approved command row.
In --loopback mode the Supabase box is an in-process fake PostgREST and the MQTT hop is
optional (--mqtt adds an embedded broker). In production the boxes are your broker and your
Supabase project — same code, same tables.
Quickstart
PowerShell 5.1+ and Python 3.10+ (the installer finds a suitable Python or tells you where to get one). From the repo root:
# install: venv + all packages + preflight (idempotent, safe to re-run)
powershell -ExecutionPolicy Bypass -File install.ps1
# start the full loopback demo — prints the console URL and opens your browser
.venv\Scripts\python.exe -m yantraops up --loopbackOr install.ps1 -Run to do both in one go. Full walkthrough — manual
venv path, MQTT demo, real Supabase, troubleshooting:
docs/RUNBOOK-WINDOWS.md.
Python 3.10+. From the repo root:
# install + start in one command
./install.sh --run
# or by hand, in a venv of your choosing
pip install -e core -e sim -e connector -e detector -e notifier -e ops
pip install -r copilot/requirements.txt
python -m yantraops up --loopbackOpen the console URL from the READY banner. python -m yantraops
doctor preflights the environment; status health-checks a
running stack.
Host networking is the supported path (the stack binds 127.0.0.1 on ephemeral
ports, so -p publishing cannot reach it):
docker build -f docker/Dockerfile -t yantrafleet .
docker run --rm --init --network host yantrafleet # Linux
# or
docker compose -f docker/docker-compose.yml up --buildOpen the console URL from the READY line. Docker Desktop
(macOS/Windows) needs its host-networking option; details and headless/CI usage
in docker/README.md.
A single small EC2 instance runs the whole demo stack; a step-by-step guide (instance size, security group, install, keeping it running) lives in deploy/aws/README.md. The short version:
# on a fresh Ubuntu instance
sudo apt-get update && sudo apt-get install -y python3-venv git
git clone <this repo> && cd yantrafleet
./install.sh
.venv/bin/python -m yantraops up --loopback --no-openThe stack binds 127.0.0.1 by design — reach the console over an SSH tunnel rather than opening ports to the internet (see the deploy guide and SECURITY.md).
FAQ
python -m yantraops up --supabase --mqtt --broker host:1883 --no-sim and
yantrabridge ingests their state; approved operator commands go back as VDA
instantActions. Robots that don't speak VDA 5050 need an adapter on
your side, but you can already import their recordings (MCAP or JSONL, with a topic
map) via python -m yantrabridge import to use the replay, incident and
analytics side of the platform.
python -m yantraops up --loopback runs a deterministic
10-robot warehouse simulation with real VDA 5050 messages, faults, missions,
incidents and maintenance findings — the full platform, entirely on your machine, no
accounts and no keys. Add --mqtt to run the same demo over a real MQTT wire.
GEMINI_API_KEY or
OPENAI_API_KEY) it answers with an LLM grounded in live fleet tools, and a
grounding guard verifies the numbers in every answer against the tool payloads; without
a key it serves offline template answers computed directly from your data — still
cited, just less conversational. Everything else on the platform is LLM-free.
commands table, with a full audit trail). Details and hardening
checklist: docs/SECURITY.md.
Scope
Things Yantrika does not do today, so you don't find out the hard way:
fleet_meta still assumes one active writer; there is no multi-region story.--loopback --mqtt on your laptop; bring your own mosquitto (or managed
broker) for anything real via --broker host:port.