# docs

VAOS™ Documentation

Everything you need to install, build, and ship vision-native agents — from a five-minute quickstart to deployment at scale.

Quickstart → GitHub ↗

VAOS™ — Documentation

VAOS is the open vision-native operating system for autonomous agents. It runs the perception → reasoning → action loop locally on your hardware, with sensor-aware confidence and validation-gated execution. Patent-pending.

New here? Start with the 5-minute quickstart, then come back here to deep-dive concepts.

Installation

VAOS supports Linux on ARM and x86. Python 3.10+ is required.

installbash
# Install VAOS Core via pip
pip install vaos

# Or pull a pre-built edge image (RK3588)
docker pull vaos/edge:2.0-rk3588

# Verify
vaos --version
# vaos 2.0 (RK3588 backend, plugin runtime v2)

Your first agent

An agent in VAOS is just a function that receives Scene objects and emits actions. Everything else — perception, sensor management, memory — is handled by the runtime.

hello_agent.pypython
from vaos import Runtime, Agent

runtime = Runtime(camera=0, hardware="auto")

async def on_scene(scene):
    if scene.event.type == "restricted_zone_entry":
        await runtime.action("notify", target="safety-team")

runtime.attach(Agent(on_scene=on_scene))
runtime.start()

Sensor Manager

The Sensor Manager continuously evaluates each input sensor's reliability — calibration drift, exposure, occlusion, thermal — and feeds that signal into inference confidence. This is one of three patent-pending core innovations of VAOS.

The manager exposes per-frame reliability scores, automatic sensor selection in multi-camera setups, and adaptive ISP control.

SignalRangeNotes
calibration0.0 – 1.0Geometric and photometric drift
exposure0.0 – 1.0Frame exposure quality
occlusion0.0 – 1.0Estimated coverage / blocked
thermal0.0 – 1.0SoC thermal headroom for inference

Memory Engine

VAOS keeps a persistent local store of scene → action → outcome triples. This lets sites learn themselves over time without any cloud dependency. The engine is local-first and supports backup, federation, and privacy-preserving sync between sites.

Agent Runtime

The agent runtime executes plans through a plan → validate → execute loop. Every recommended action is validated against current scene confidence before execution. Patent-pending.

  • Plan — propose action from agent logic
  • Validate — gate against scene confidence and sensor reliability
  • Execute — emit to actuator / API / dashboard
  • Record — capture into memory engine

Deployment

VAOS deploys as a single binary with optional plugin runtime. Production builds support OTA updates, A/B rollouts, and node federation. See downloads for pre-built images.

Hardware support

  • Rockchip RK3588 / RK3688 + RK1828
  • NVIDIA Jetson (Orin family)
  • Raspberry Pi 5 + AI accelerators
  • NXP i.MX 8M Plus
  • x86 with discrete GPU (development)

OTA updates

VAOS supports signed OTA updates of the runtime, agents, and plugins. Federation supports A/B rollouts, canary nodes, and rollback.

Observability

The runtime exposes structured logs, Prometheus metrics, and a live MQTT trace. The developer dashboard consumes these for debugging and validation.

Patent pending. VAOS includes patent-pending innovations in sensor-aware perception, scene contract validation, and validation-gated action runtime. Use under Apache 2.0 for the open-source components; commercial licensing available for derivative hardware/silicon products. Contact eye3.ai →