Contact Us
SOFTWARE PLATFORM · IN-VEHICLE AI

AIOS — In-Vehicle Multimodal AI Assistant

Voice, vision and gestures processed onboard — understanding, decision-making, execution and memory all happen in the vehicle. Edge-first, delivered by feature, security built in: fully functional offline, raw data never leaves the car.

0.5–7B
Onboard multimodal models · tuned and customized
0
Network dependency for voice interaction · fully functional offline
3 tiers
Memory architecture · in-vehicle tiered storage

One assistant covering every in-vehicle interaction

AIOS is delivered as an integrated assistant rather than a collection of features: unified onboard multimodality, a full in-vehicle pipeline, and feature points as the unit of delivery and acceptance. Tool calls, intent understanding and customer signal conventions are baked into the model during training instead of inferred at runtime — a new feature takes one round of model tuning.

Core Mechanisms

Edge-first · Security built in · Integrate, never replace

01

Edge-First

Understanding, decision-making, execution and memory happen onboard, with core capabilities independent of the cloud; consistent experience in garages and tunnels, privacy by physical isolation.

02

Dual-Path Routing

Vehicle-control requests follow deterministic rules straight to SOA services with millisecond responses and controllable results; open-domain dialogue and multimodal requests use onboard generative models (SLM / VLM).

03

Safety Fence

AI may propose, but vehicle-control commands must pass multi-dimensional safety assessment (physical constraints / regulations / occupant safety / scenario risk); high-risk operations require confirmation and everything is auditable.

04

Integrate, Never Replace

Standard interfaces fit into existing HMI, perception stacks, cloud and OTA systems; control commands never come from generative models, and signal changes are configuration-level only.

05

Feature-Point Delivery

You define the feature list; we own tuning, deployment and vehicle integration end to end — one consistent yardstick for procurement, implementation and acceptance.

06

Six Feature Categories

Voice interaction, visual perception, gesture control, identity and personalization, proactive services and sentry review — combinable per list and extensible.

Technical Specifications

Model Scale
0.5–7B multimodal models; tool calls and domain knowledge fixed during training, supporting fine-tuning and distillation
Deployment
Edge-first, running in a Linux environment; when the cockpit domain runs Android, it is hosted in an in-domain Linux container with resource isolation
Vehicle-Control Path
Deterministic NLU straight to SOA services, bypassing generative models; fast responses, controllable and verifiable results
Interaction
See-and-say, multi-turn context, cross-app task orchestration, screen content understanding (VLM)
Memory
Session layer ephemeral · long-term private layer encrypted onboard · portable layer desensitized; viewable, deletable, memory never leaves the car
Integration Interfaces
HMI via MCP / REST / WebSocket; vehicle signals via VHAL / SOME-IP / DDS; OTA via DoIP / UDS
Security
Deterministic vehicle control plus safety fence; raw perception data stays onboard, cloud queries are desensitized first; full execution auditability
Continuous Evolution
Continuous OTA updates for features and models; the same feature adapts across compute platforms

Experience

Voice Interaction

Fast and reliable for vehicle control, open for conversation: wake-up, recognition, understanding and synthesis run onboard — works in garages and tunnels.

Proactive Services

Proactive suggestions based on fatigue, child presence, battery and calendar states, executed through the safety fence — proactive without being intrusive.

Personalization & Emotion

Multi-user profiles with fused face and voice recognition, preference memory and emotional cockpit orchestration — a car that knows you better over time.

Related Products & Solutions