RUYi Dynamics · RY OS

RY OS for Real Robots

RY OS separates human-like task intelligence from realtime robot execution. The agentic layer understands, remembers, plans, and generates skills; the runtime layer validates, schedules, supervises, and executes through a typed safety boundary.

18event types

11decoupled agents

4skill types

5task priorities

Unified RY OS Architecture

01 · Agentic Intelligence

Agentic Intelligence

Voice, digital channels, text, gesture and visual context become plans, memory calls, tool use and skill dispatch.

interaction loop

planner + memory

skill engine

02 · ryos-ar Boundary

ryos-ar

Typed commands, state, events, heartbeat, safety and e-stop contracts.

03 · Robot Realtime Execution

Robot Realtime Execution

Runtime-owned validation, scheduling, safety, process supervision, controllers, watchdogs and actuators.

standalone runtime

zero-copy comms

safety + recovery

012345678901234567890123456789012345678901234567890012345678901234567890123456789012345678901234567890123456789012345678901

From multimodal intent to grounded robot tasks

Voice, digital channels, text, gesture, body-language, touch, and visual context are fused into an interaction turn. The agentic brain then uses dialog, memory, tools, permissions, body capability, and policy reasoning to decide what should happen.

0123456789012345678901234567890123456789012345678900123456789012345678901234567890123456789012345678901234567890123456789012

A narrow typed contract between intent and execution

RY OS does not send natural language directly to robot control. The shared ryos-ar contract turns intent into command, state query, cancel, safety, and e-stop schemas, with state, heartbeat, acknowledgement, rejection, progress, fault, and completion events flowing back.

01234567890123456789012345678901234567890123456789001234567890123456789012345678901234567890123456789012345678901234567890123

Stable physical execution with safety-owned runtime control

The realtime runtime owns device lifecycle, process supervision, realtime scheduling, communication, execution state machines, watchdogs, controller arbitration, safety gates, and actuator ownership. Agentic intelligence may request execution; it does not own motors or hard realtime loops.

RY OS · Execution Boundary

Unified Agent-Runtime Flow

A single request crosses two layers without blurring ownership: agentic cognition decides intent; robot realtime execution owns validation and physical actuation.

Agentic Intelligence Detail

Input Fusion, Reasoning, Skill Output

Human Signals

multi-modal input

Voice

turn-taking, barge-in

Digital Channel

task request, context

Gesture

pointing, selection

Body / Touch

social cues, contact

Agentic Brain

reason before acting

Interaction Manager

clock, interrupt, fusion

Intent + Context

dialog, memory, scene state

Planner

decompose and choose tools

Skill Output

tools and robot requests

Use Tool

query, search, operate

Dispatch Skill

known robot behavior

Generate Skill

new scripted behavior

Typed Request

send through boundary

Robot Realtime Execution Detail

Typed Boundary, Realtime Core, Device Execution

Boundary Contract

agent-facing interface

Command

typed intent from agent

State

observable robot facts

Heartbeat

liveness and timeout

Event

result, fault, progress

Realtime Core

stable system base

Zero-copy Transport

high-rate state path

Process Supervisor

start, stop, recover

Realtime Scheduler

timing and priority

Safety Gate

watchdog, arbitration, e-stop

Robot Devices

physical execution

Sensors

camera, lidar, proprioception

Controllers

motion and behavior control

Actuators

physical robot output

01 · Agentic Intelligence

From multimodal intent to grounded robot tasks

Realtime Interaction Loop

VAD, barge-in, partial text, audio queue, cancellation timing, and multimodal fusion.

Planner + Memory

Task decomposition, world memory, spatial memory, procedural memory, and reflection.

Skill Engine

Existing skills, generated Python skills, tool calls, MCP calls, and body-requirement gating.

02 · ryos-ar Boundary

A narrow typed contract between intent and execution

Command API

Typed command transport for action requests, cancellation, safety, and emergency stop.

State + Event Stream

Robot state, observations, device state, telemetry, safety events, and heartbeat.

Contract Tests

Generated schemas and bindings keep agent code and runtime code aligned.

03 · Robot Realtime Execution

Stable physical execution with safety-owned runtime control

Standalone Runtime

C++ execution base with process model, lifecycle supervision, logging, and production control.

Realtime Communication

Low-latency middleware, zero-copy data paths, telemetry, record/replay, and visualization.

Safety + Recovery

Validation, arbitration, watchdogs, e-stop, rejection, recovery, and fault reporting.

Unified request flow

Multimodal input

Realtime fusion

Intent + policy

Skill dispatch

ryos-ar typed command

Runtime validation

Controller execution

Events + learning

Swap the body. Keep the brain.

RY OS lets the intelligence layer move across robot bodies while keeping physical actuation governed by a stable, observable, safety-owned runtime.