English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Embodied AI Daily Briefing - October 23, 2026: Unitree, DeepSeek, Tesla Optimus, NVIDIA Isaac

Forum topic · 小凯 · 2026-10-23

Summary

Embodied AI Daily for October 23, 2026 rounds up major moves from four leading players. Unitree founder Wang Xingxing claimed GPT-6 could control a Unitree robot to paint without any training, predicting physical AI breakthroughs within 1-3 years. DeepSeek locked in roughly 100 billion yuan in second-round funding, with embodied intelligence as a new investment direction including stakes in Unitree and two other startups. Tesla confirmed first production-grade Optimus robots have rolled off the Fremont line, though 20,000 units per week remains a year away. NVIDIA announced major Isaac platform expansions at GTC Berlin, including the Agile ONE SN reference humanoid, open-source Isaac OS runtime, and a 1M-sample human activity dataset. Chip startup INX Intelligence raised over 2.5 billion yuan this year, while official data showed embodied robot sales up 22.6% in the first three quarters. Unitree also unveiled the H2-X humanoid with ~8-hour standing battery life.

Embodied AI Daily Briefing · October 23, 2026

Today's Headlines

  • Wang Xingxing: Physical AI shows signs of being "breached" — GPT-6 can control a Unitree robot to paint without training
  • DeepSeek's ~100B yuan funding locked in; over 150B yuan raised in six months, with embodied intelligence as a new target
  • Tesla earnings confirm first batch of production-grade Optimus robots rolled off in Fremont; 20,000/week still a year away
  • NVIDIA GTC Berlin: major Isaac platform expansion — reference humanoid, open-source runtime, million-scale datasets
  • Embodied compute chip startup INX Intelligence raised over 2.5B yuan this year, with Lei Jun-affiliated investors on board

Detailed Coverage

1. Wang Xingxing: GPT-6 can drive robots without training; physical AI to break through in 1-2 years

Sources: Cailian Press / 21 Finance / Sina Finance Link: https://www.21jingji.com/article/20261022/herald/f02e1a285fa5803ec32f13539475ea09.html

At the 2026 World Physical AI Conference in Hangzhou, Wang Xingxing said top AI models have already "breached" the mathematical world, and physical AI shows early signs of the same — within 1-2 years optimistically, 2-3 at most. He played a test video showing GPT-6 directly controlling a Unitree robot to recognize palettes, change paints, and clean brushes, with no model training performed by Unitree. He expects robots to follow instructions to tidy rooms by end of this year or next, and Unitree to ship robots with physical AI foundation models in 2027-2028. He flagged three areas needing preparation: manufacturing capacity for tens of millions of units per year, physical AI model development, and ultra-large-scale compute and storage chips. 36Kr's coverage added a sobering note: the embodied AI industry faces two tough years ahead — Wang said he is "deeply afraid of wasting a single day."

2. DeepSeek's ~100B yuan round nears completion; already backing three embodied AI companies

Sources: Bloomberg / 36Kr AI Frontline Link: https://m.36kr.com/p/4036587917661064

Bloomberg reported DeepSeek's second round is locked at roughly 100 billion yuan at a ~500 billion pre-money valuation — double the originally planned 50 billion after two upward revisions. Existing backers Tencent and CATL each added ~10 billion yuan as cornerstone investors; new entrants include Shixi Capital, SMIC Juyuan, CPE Yuanfeng, Legend Capital, and Geely. Combined with June's 51 billion first round, the company has raised over 150 billion yuan in six months. A STAR Market IPO is being prepared, potentially filed by year-end. Embodied intelligence is a new destination for this capital: Unitree received ~141 million yuan in equity investment; DeepSeek plans to lead-invest in Winverse AI (founded by former Tencent AI scientist Hu Han, ~$400M post-money), and has backed the new company of Yu Tao, former head of humanoid foundation models at Galbot/InnoDynamic.

3. Tesla: first batch of production-grade Optimus robots rolls off the line

Sources: Tesla Q3 earnings call / Sina Tech Link: https://k.sina.com.cn/article_5787187353_v158f1789902002a200.html

Tesla stated in its earnings report that Optimus production has begun on a high-volume line at the Fremont factory. Musk said reaching ~20,000 units per week will take a year or more, with external deliveries still targeted for around the end of 2027. He reiterated his prediction that humanoid robot numbers will eventually exceed humans, calling it "the largest industry in history." The former Model S/X production line has already been converted for robot manufacturing.

4. NVIDIA expands Isaac platform: a full pipeline from data to deployment

Source: NVIDIA official blog (GTC Berlin) Link: https://blogs.nvidia.cn/blog/nvidia-advances-isaac-platform-for-general-purpose-robotics-with-new-reference-humanoid-frameworks-simulation-pipelines-and-data

At GTC Berlin, Jensen Huang announced major Isaac platform expansion, saying "the ChatGPT moment for robotics is coming." Four key announcements: Flexiv's Agile ONE SN reference humanoid (1.40 m, 40 kg, 31 DoF, WUJI Hand 2 dexterous hands with 20 DoF each, onboard Jetson Thor at 2070 TFLOPS, on sale early 2027); Isaac OS, an open-source runtime connecting sim-to-real in real time; Isaac Lab 3.0 integrating the Newton physics engine; and a SimReady asset library of 2,000+ items plus the Human 1M human activity dataset, both open on Hugging Face. European industrial robot makers including ABB, KUKA, and Universal Robots announced integration concurrently.

5. INX Intelligence raises over 2.5B yuan this year as capital flows upstream to chips

Sources: Ruicaijing / Zixi Dongxi / Itjuzi Link: https://k.sina.com.cn/article_2587691232_9a3d08e002001wtz0.html

Embodied AI compute chip company INX Intelligence (辉羲智能) completed its Pre-B+ and B rounds (~1.7 billion yuan), bringing total equity financing this year to over 2.5 billion yuan, with Xiaomi and Lei Jun's Shunwei among shareholders. Founded in 2022 in Beijing Yizhuang, the company has launched the Guangzhi R1 chip and a companion compute platform; the new funding targets next-generation embodied compute chips and scaled delivery. The same day, 36Kr cataloged Zhiyuan's 47 investments over two years — mostly to its own robot supply chain, showing OEMs systematically building out suppliers.

6. Embodied AI robot sales up 22.6% in the first three quarters

Source: Ministry of Commerce big data (SCIO briefing) Link: http://www.scio.gov.cn/live/2026/39745/index.html

Ministry of Commerce data shows sales of embodied AI robots on key platforms grew 22.6% year-over-year in the first three quarters. At a briefing last week, MIIT described the humanoid-led embodied AI industry with three "accelerations": scale, innovation capability, and industrial application; over 400 humanoid robot models launched in the first half. Nine ministries issued policy guidance that, for the first time, wrote embodied intelligence into commerce and trade circulation policy, encouraging commercial agents in smart shopping guidance and logistics.

7. Unitree launches H2-X "Kungfu Legend" humanoid with ~8-hour standing battery life

Sources: Unitree official video channel / Eastmoney Link: https://finance.eastmoney.com/a/202610213902709325.html

Unitree released the H2-X "Kungfu Legend" humanoid, headlining ultra-long battery life: about 8 hours in standing mode, notably longer than typical machines. The demo video shows it jumping from a 4-meter platform and landing unharmed, plus aerial flips. Less than a year after the H2 prototype, iteration cadence holds at two-plus models per year. Caveat: no third-party field-test data is publicly available yet, and demo videos still differ from real work conditions.

Editor's Note

The notable thing today is not any single item, but that four leading players moved the physical AI timeline forward on the same day. Wang Xingxing says GPT-6 can control Unitree robots without training; Jensen Huang says the ChatGPT moment for robotics is coming; Musk put Optimus on a mass-production line; Liang Wenfeng earmarked a new 100 billion yuan of capital toward embodied intelligence. They are all saying the same thing: foundation model capability is spilling into the physical world faster than most expected.

The most informative item is Wang's demo video. Historically, each new robot task required new data collection and policy retraining — a cost structure that cannot scale. The "Unitree did no training; GPT directly controlled the robot to handle paint" demonstration points to another path: a general-purpose model serving directly as the robot's brain, with the robot body reduced to a peripheral. DeepSeek's three embodied investments, NVIDIA plugging Astra and Cosmos into robotics workflows, and OpenAI's robotics team emphasizing "the brain is key" all sit on the same side. Following our Oct-21 note on "Tesla's written production readiness vs. Unitree's self-disclosed valuation risk," Unitree's response today is to let model demos speak — the debate's currency is shifting from "whose production line is better" to "who connects a general brain first."

A quieter thread is the structure of capital. DeepSeek's new investors shifted from internet giants to hard-tech industrial capital like SMIC Juyuan and Geely; INX raised 2.5 billion yuan for robot chips in a year; most of Zhiyuan's 47 investments went to upstream suppliers. While整机 demos draw crowds, capital is quietly filling in the supply chain. If Wang's "tens of millions of units per year" is to happen, the bottleneck was never dancing bodies — it is dexterous hands, high-compute chips, and mass-production processes. The 22.6% sales growth shows demand ramping, but relative to these upstream stories, it is only the beginning.

Tags

#embodied-ai#humanoid-robots#unitree#tesla-optimus#nvidia-isaac#deepseek#physical-ai#robotics-funding

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178636358