「豆包汽车」落地:豆包座舱助手随荣威家越07开启预售,AI智能体跨终端布局提速Bean Bag Car” lands: Bean Bag Cockpit Assistant opens pre-sale with Roewe Family Crossing 07, AI agent cross-terminal layout accelerates


AI free token/AI免费积分

“豆包汽车”落地,AI智能体跨终端布局提速

问AI · 相比手机跨应用难汽车智能体为何易闭环?

近日,豆包与火山引擎联合构建的AI原生车载助手“豆包座舱助手”正式发布。首款搭载车型荣威家越07将于9月21日启动预售,上汽奥迪相关车型年内亮相。就在此前,豆包手机助手消费者版本发布,努比亚NaviX Ultra以“全球首款AI智能体手机”身份登场。

手机端与车载端的接连落子,释放出一个清晰的信号:AI智能体正从手机向汽车等硬件终端延伸。华为鸿蒙座舱的小艺智能体已打通导航、控车、生活服务等情形,小米的超级小爱也在从“对话机器人”升级为具备执行能力的智能体。跨终端布局,正在成为行业共同方向。

豆包此前已进入汽车,这次的不同在哪?

豆包并非首次进入汽车。今年7月31日起,特斯拉中国开始分批次推送2026.14.13版本车机软件,涵盖Model 3、Model Y、Model S、Model X全系车型,核心升级内容就是引入豆包大模型。火山引擎今年4月公布的数据表明,搭载豆包大模型的智能汽车已超过700万辆,覆盖50多个汽车品牌、145个车型。

但在此之前,豆包主要以底层模型或语音能力的形式存在。使用者接触到的往往是车企自有的助手名称和形象,能力集中在实时信息查询、自然对话、唱歌讲故事。在特斯拉车上,豆包以独立App形态存在,不介入底层车辆控制权限,涉及行车安全的执行指令仍沿用原有通道。

这次,家越07上的豆包座舱助手直接面向使用者,使用者叫的是豆包,看到的也是豆包。它不再只是“听懂人话”,而是参与“感知—决策—执行”闭环。官方演示中,一句“午睡一会,1点叫我”,豆包座舱助手会自动按照使用者习惯调整座椅、关闭遮光板、调低风量、播放轻音乐、启动按摩,到点准时叫醒,并提醒下午日程。能做到这些,靠的是底层能力的打通。荣威累计投入数百亿资金构建整车智能电子电气架构,构建了2000+SOA全域服务接口。座椅、空调、车窗等硬件被原子化解耦,AI可以自由调用。

当下,家越07展车已陆续抵达全国多家荣威授权经销商门店。此前展厅里的展车,车载AI能力还未解锁,厂家完成远程OTA升级之后,豆包车载AI已经可以正常感受。有意向车主说,“以前用语音助手像在跟机器下命令,现在更像个副驾在旁边帮你操心。”不过他也坦言,手机上的豆包已经用顺手了,“如果车里的豆包跟手机上不是同一个,那还得重新适应。”

汽车比手机更合适Agent先行落地吗?

当AI智能体开始跨终端布局,汽车为什么成了关键战场?

手机上的智能体面临一个现实难题:模型已经会干活了,可App不一定让它干。想帮你点外卖、叫车、订酒店,意味着要进入几十个第三方应用,涉及权限、安全、协议和商业利益的多重博弈。而汽车里,空调、座椅、车窗、导航大量能力本就集成在一套整车平台中。只要车企愿意开放底层接口,AI不需要跨应用去“找权限”。

汽车端的长处还不止于此。它天然有大量传感器和执行器,定位、摄像头、麦克风、座椅状态构成AI的“眼睛”和“耳朵”,空调、车窗、音响、灯光则成为AI的“手”和“脚”。过去的汽车不缺这些,缺的是一个能把它们串起来的大脑。

但要让AI真正执行任务,前提是知道哪些事不能做。荣威为AI设定了黑区、灰区、彩区三类权限:涉及行驶安全的能力最高权限锁定,满足安全条件时允许有限控制,不影响安全时可完全自主调度。刹车、转向等核心驾驶控制,不向座舱大模型开放。如何在“放权”与“收权”之间找到平衡,是整个行业都在摸索的课题。

更现实的挑战在商业逻辑层面。车企长期习惯按BOM(整车零部件成本)来核算一切,每个零部件装上车就不再产生额外支出。但座舱AI推理不同,它的费用随使用者使用量持续增长。若仍按传统的固定物料成本来看待推理费用,逻辑上就会浮现能力越受欢迎、使用者调用量越大、企业反而越亏损的困局。在业内人士看来,算力不应被视为研发成本,而是直接支撑使用者感受和业务流的核心生产资源。

业内人士指出,2026年是“智能体之年”,AI智能体跨越全品类终端,以一种“计算连续体”的形式给出持续服务,汽车正演进为智能体AI最关键的移动载体之一。这并不意味着汽车端的落地没有门槛。从语音识别到意图理解,从单次指令到长期记忆,从回复信息到真正打通服务生态完成闭环,每一步都需要大量工程化投入。

记者 吴茜 王柳 责任编辑 罗皓 编辑 刘春春 审核 何晓蓉

本文来源:https://news.qq.com/rain/a/20260920A06GGI00
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

The arrival of "Bean Bag Car," AI agent cross-terminal layout accelerated

头像

Jinguang News

Published by the official account of Sichuan Chengdu Daily "Jinguang News" on September 20, 2026, at 11:57

Asking AI. Why is it easier for car agents to close the loop than for mobile phones to cross apps?

Recently, the AI-native vehicle assistant "Bean Bag Cabin Assistant" jointly built by Bean Bag and Volcanic Engine was officially released. The first model to feature it, Roewe Jia Yue 07, will begin presale on September 21, with relevant Audi models to be unveiled within the year. Prior to this, the consumer version of the Bean Bag Mobile Assistant was released, and Nubia NaviX Ultra debuted as the "world's first AI agent phone."

The consecutive moves on the mobile and vehicle fronts send a clear signal: AI agents are extending from mobile phones to other hardware terminals like cars. Huawei Harmony Cabin's Xiaoyi agent has already bridged navigation, vehicle control, and lifestyle services, and Xiaomi's Super Xiaoai is upgrading from a "dialogue robot" to an agent with execution capabilities. Cross-terminal layout is becoming the common direction for the industry.

Bean Bag has entered cars before. What is different this time?

Bean Bag is not new to cars. Starting July 31 this year, Tesla China began rolling out version 2026.14.13 of its vehicle software to all Model 3, Model Y, Model S, and Model X vehicles, with the core upgrade being the introduction of the Bean Bag large model. Data from Volcanic Engine in April shows that over 7 million smart cars are equipped with the Bean Bag model, covering 50+ brands and 145 models.

However, before now, Bean Bag was mainly present as an underlying model or voice capability. Users interacted with the automaker's own assistant names and images, focused on real-time queries, natural dialogue, music, and storytelling. In Tesla cars, Bean Bag exists as a separate app without access to underlying vehicle controls; safety-critical commands still flow through the original channel.

This time, the Bean Bag Cabin Assistant on Jia Yue 07 faces the user directly; users call for and see Bean Bag. It no longer just "understands speech" but participates in the "perception-decision-execution" loop. In the demo, a single command like "Take a nap, wake me at 1 PM" triggers a cascade: the assistant adjusts seats, closes blinds, lowers fan speed, plays soft music, activates massage, and wakes the user precisely, reminding them of afternoon plans. This is possible because underlying capabilities are now integrated. Roewe invested hundreds of billions in its intelligent electronic-electrical architecture, creating 2000+ SOA services that atomize hardware like seats and climate control, allowing the AI to call them freely.

As of now, many Roewe dealerships nationwide have received the Jia Yue 07 demo. Previously, the car's AI was locked, but after a remote OTA update, the Bean Bag AI is now active. A prospective buyer notes, "Before, talking to the assistant felt like commanding a machine; now it feels like a copilot looking out for me." He adds, though, that he's already used to Bean Bag on his phone — inconsistent experiences between devices would be unwelcome.

Are cars more suitable for agent landing than phones?

As AI agents go cross-terminal, why have cars become a key battleground?

Mobile agents face a fundamental problem: the model can do the work, but the apps may not let it. Ordering food, booking rides, or reserving hotels requires access to dozens of third-party apps, involving permissions, security, and commercial interests. In cars, however, climate, seats, windows, and navigation are already part of a single platform. If the automaker opens the APIs, the AI doesn't need to cross apps to gain permission.

Cars also boast abundant sensors and actuators — location, cameras, mics, seat sensors act as the AI's eyes and ears, while climate, windows, audio, and lights are its hands and feet. The missing piece has been a brain to coordinate them.

For true execution, the AI must know what *not* to do. Roewe sets black, grey, and color zones — locking out safety-critical functions, limiting them under safe conditions, and fully enabling non-critical ones. Braking and steering are off-limits to cabin models. Balancing power and restraint is an industry-wide challenge.

There's also a commercial-logic hurdle. Automakers traditionally account for costs via BOM (bill of materials), where parts are fixed-cost. But AI inference scales with usage — more use = more cost. If treated like physical BOM, popularity leads to losses. Industry experts argue that compute should be seen as operational overhead supporting experience, not R&D expense.

2026 is the "Year of the Agent," says an insider, with AI forming a continuous compute fabric across devices, cars becoming key mobile carriers. But this isn't trivial — from speech recognition to intent understanding, long-term memory, and full service integration, each step requires heavy engineering.

Reporters: Wu Qian, Wang Liu | Editor-in-Chief: Luo Hao | Editor: Liu Chunchun | Reviewer: He Xiaorong

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry