中國機器人製造商已掌握動作，接下來是大腦競賽

作者：蘿莉・陳與愛德華多・巴普蒂斯塔By Laurie Chen and Eduardo Baptista

北京，11月2日（路透社）—— 經過多年打造能跑步、跳舞和後空翻的人形機器人後，中國企業正把資源投入另一項挑戰：開發能讓這些機器在更廣泛工作中派上用場的 AI。BEIJING, Nov 2 (Reuters) - After years of building humanoid robots capable of running, dancing and performing backflips, Chinese companies are pouring resources into a different challenge: developing the AI that could make those machines useful across a wider range of jobs.

這項轉向正帶動具身人工智慧（embodied AI）投資熱潮。這個領域聚焦於讓機器人理解周遭環境、學習陌生任務，並在不必為每項工作個別編寫程式的情況下自行適應的軟體。The shift is fuelling an investment boom in embodied AI, a field focused on software that allows robots to understand their surroundings, learn unfamiliar tasks and adapt without having to be programmed for every job.

資料研究公司 IT桔子的數據顯示，2026年上半年，中國具身 AI 企業透過322筆交易募得935億元人民幣（130億美元），是前一年同期的5倍。資金除了流向製造商，也越來越多投入專門開發機器人智慧的業者。Chinese embodied AI companies raised 93.5 billion yuan ($13 billion) across 322 deals in the first half of 2026, five times the amount a year earlier, according to data research firm ITjuzi. Funding is increasingly reaching specialist developers of robot intelligence alongside manufacturers.

總部位於上海的科技顧問公司 ZenGen Labs 創辦人德莫特・麥格拉斯（Dermot McGrath）表示，隨著競爭壓低硬體價格，機器人製造商已「得出結論，利潤不會來自硬體，而是來自大腦」。As competition drives down hardware prices, robot makers have "concluded that the margin won't come from the hardware but from the brain," said Dermot McGrath, founder of Shanghai-based technology consultancy ZenGen Labs.

這個新興產業正在美中科技競爭中開闢新戰線。路透社訪問5家中國機器人公司的創辦人和高階主管，以及產業分析師後得知，美國企業在部分先進機器人 AI 系統上領先，而中國開發商希望憑藉製造實力和取得真實世界訓練資料的能力，挑戰美國競爭對手。The emerging sector is opening a new front in US-China tech rivalry. American companies lead in some advanced robot AI systems, while Chinese developers hope their manufacturing strength and access to real-world training data will help them challenge their US competitors, according to Reuters interviews with founders and top executives at five Chinese robotics firms, as well as industry analysts.

哪個國家最能把 AI 與機器結合，就可能在工廠與物流自動化上取得優勢，進而可能重塑工業生產力和全球機器人技術市場。The country that best combines AI with machines could gain an advantage in automating factories and logistics, potentially reshaping industrial productivity and the global market for robotics technology.

Spirit AI 共同創辦人高陽（Gao Yang）表示：「大腦確實是完整機器人技術堆疊中最薄弱的一環。」"The brain is indeed the weakest link in the complete robotics stack," said Spirit AI co-founder Gao Yang.

智慧競賽RACE FOR INTELLIGENCE

包括 Physical Intelligence、Skild AI 和 Generalist 在內的美國新創公司，正在開發旨在讓機器人無須大量重新訓練就能學會不同技能的軟體。American startups, including Physical Intelligence, Skild AI and Generalist, are developing software intended to teach robots different skills without extensive retraining.

Generalist 表示，其 GEN-1.5 模型在觀看長度介於3到12秒的示範後，完成約60%的簡單任務。中國的 X Square 已發表研究，顯示機器人能從一段人類影片中習得操作技能，而不必針對每項任務重新訓練模型。Generalist says its GEN-1.5 model completed around 60% of simple tasks after viewing a demonstration lasting between 3 and 12 seconds. China's X Square has published research showing robots acquiring handling skills from a single human video without retraining the model for each task.

機器人顧問格奧爾格・施蒂勒（Georg Stieler）表示：「領先的美國實驗室正在設下很高的標竿，尤其是在機器人能多快學會並重複使用技能方面。」他認為中國最強的新創公司是可信的競爭者。"The leading American labs are setting a demanding benchmark, particularly in how quickly robots can acquire and reuse skills," said robotics consultant Georg Stieler, who considers China's strongest startups credible competitors.

美國開發商受益於先進 AI 研究，以及更容易取得強大運算晶片；中國企業則能與機器人製造商和工業客戶密切合作，測試並改進其系統。While US developers benefit from advanced AI research and greater access to powerful computing chips, Chinese companies can work closely with robot manufacturers and industrial customers to test and improve their systems.

公司揭露資料顯示，包括 X Square、Galaxea 和 Spirit AI 在內的中國專業新創公司，估值各自已達約30億美元。Specialist Chinese startups, including X Square, Galaxea and Spirit AI, have each reached valuations of roughly $3 billion, according to company disclosures.

X Square 共同創辦人兼營運長楊謙（Yang Qian）表示：「我們希望機器人從由程式碼驅動，轉向由常識驅動。」"We want robots to shift from being driven by code to being driven by common sense," said X Square co-founder and COO Yang Qian.

既有製造商也在投資。香港上市的優必選（UBTech）刊登首席科學家職缺，負責 AI 模型開發，薪酬最高可達1.24億元人民幣（1800萬美元），包含股權和福利。一名聽取過宇樹科技簡報的人士表示，總部位於杭州的宇樹科技是全球出貨量最大的人形機器人製造商之一，約有30名研究人員正在開發可預測物體行為的 AI。Established manufacturers are investing too. Hong Kong-listed UBTech advertised a chief scientist position overseeing AI model development with compensation of up to 124 million yuan ($18 million), including equity and benefits. Hangzhou-based Unitree, one of the world's largest humanoid robot makers by shipments, has about 30 researchers developing AI that predicts how objects behave, according to a person briefed by the company.

宇樹科技未回應採訪請求。Unitree did not respond to a request for comment.

訓練競賽THE TRAINING RACE

與從大量網路文字中學習的聊天機器人不同，機器人需要實體互動的範例，從抓握物體到處理柔性電線皆然。Unlike chatbots, which learn from vast quantities of internet text, robots need examples of physical interactions, from gripping objects to handling flexible wires.

中國企業正以不同方式蒐集這些資訊，有些雇用人員示範任務，有些使用電腦模擬，也有些兩者並用。這個領域仍相當擁擠，尚無公司確立明確領先地位。Chinese companies are pursuing different approaches to collecting that information, with some employing people to demonstrate tasks, others using computer simulations, and some combining both. The field remains crowded, with no company establishing itself as the clear leader.

高陽表示，Spirit AI 已聘用約1000人，記錄日常與工業活動。ACE Robotics 則使用穿戴式感測器，在商店和飯店捕捉工作人員的動作和握力。Spirit AI has hired around 1,000 people to record everyday and industrial activities, Gao said. ACE Robotics uses wearable sensors to capture workers' movements and grip forces in stores and hotels.

螞蟻集團旗下機器人事業 Robbyant 表示，該公司使用16萬小時素材訓練模型，其中包括5萬小時由人類引導的機器人動作。該公司也使用來自17個品牌、20種機器人配置的資料訓練軟體，目標是在不同機器上使用同一套 AI。Ant Group's robotics unit Robbyant said it trained a model using 160,000 hours of material, including 50,000 hours of human-guided robot movements. It also trained software on 20 robot configurations from 17 brands, aiming to use the same AI across different machines.

DexForce 採取另一種做法，利用電腦生成環境來訓練機器人。營運長妮可・楊（Nicole Yang）表示，該公司能在6到8小時內讓部分工業任務準備部署，並回報在特定精密操作中的成功率超過99.9%。DexForce is taking another approach, using computer-generated environments to teach robots. Chief operating officer Nicole Yang said the company can prepare some industrial tasks for deployment within 6 to 8 hours and reported success rates above 99.9% in specified precision operations.

但為 AI 訓練設計模擬環境的空間智慧公司 Manycore Tech 的傑瑞・王（Jerry Wang）表示，模擬無法完美重現實體互動。But simulations cannot perfectly reproduce physical interactions, said Jerry Wang of spatial intelligence company Manycore Tech, which designs simulated environments for AI training.

X Square 的楊謙表示：「資料品質比資料多樣性更重要，也比單純大規模蒐集資料更重要。」"Data quality is more important than data diversity, and more important than simply collecting data at massive scale," X Square's Yang said.

中國機器人公司也面臨其他逆風，包括美國出口管制導致先進 Nvidia 晶片短缺，以及監管機關對估值膨脹和產業過度擁擠的疑慮，使 IPO 融資受到擠壓。Chinese robot firms also face additional headwinds, including shortages of advanced Nvidia chips due to US export controls, and a squeeze on IPO funding amid regulatory concerns over inflated valuations and overcrowding.

負責開發 JoyIn Technology 的 Aether 機器人 AI 的朱明軒（Zhu Mingxuan）表示，隨著機器人功能越來越強大，先進晶片短缺將成為中國具身 AI 產業的嚴重問題。As robots become more powerful, the shortage of advanced chips will become a serious problem for China's embodied AI industry, said Zhu Mingxuan, who leads development of JoyIn Technology's Aether robot AI.

從訓練走向工作FROM TRAINING TO WORK

JoyIn 和其他開發商已在測試其軟體是否能讓機器人處理陌生問題。JoyIn and other developers are already testing whether their software can enable robots to tackle unfamiliar problems.

朱明軒描述一項實驗：一台機器人一再無法把耳機放進超出其觸及範圍的櫥櫃。Zhu described an experiment in which a robot repeatedly failed to place headphones in a cupboard beyond its reach.

朱明軒表示，後來工程師向它展示人類如何完成這項任務的影片，這台機器人便把一個箱子推向櫥櫃，並試著踩上去。Then engineers showed it videos of how people managed the task, and the robot pushed a box towards the cupboard and tried stepping onto it, Zhu said.

她說：「它每一次都在探索不同的行為。」"Every time it was exploring a different behaviour," she said.

在中國的真實場域部署中，機器人沒有即興發揮的空間。電池製造商寧德時代（CATL）表示，Spirit AI 機器人在其中州廠進行電池組測試時連接柔性電線，插入成功率超過99%。Robbyant 表示，其機器人在上海兩家藥局處理訂單，藥品出貨前由人員檢查。Real-world deployments in China leave no room for improvisation. Battery maker CATL said Spirit AI robots connect flexible wires during battery-pack testing at its Zhongzhou plant, reporting insertion success above 99%. Robbyant said its robots handle orders in two Shanghai pharmacies, with staff checking medicines before dispatch.

但較長期的獎賞，是能在不需大量調整的情況下學會新工作的軟體。成功將取決於同一套 AI 在不同客戶和工作場所中的表現，而這將要求企業用先進晶片訓練越來越強大的模型。But the longer-term prize is software capable of learning new jobs without extensive adaptation. Success will depend how the same AI performs across different customers and workplaces and that will require companies to train increasingly powerful models with advanced chips.

中國企業正試圖透過更高效率的模型、合成資料和模擬來降低運算需求；同時，包括華為在內的國內晶片製造商也在開發 Nvidia 硬體的替代方案。目前仍不清楚這些努力能在多大程度上抵消短缺。Chinese companies are trying to reduce their computing requirements through more efficient models, synthetic data and simulations, while domestic chipmakers, including Huawei, are developing alternatives to Nvidia hardware. It remains unclear how far those efforts can offset the shortage.

「ChatGPT 時刻」'CHATGPT MOMENT'

宇樹科技創辦人王興興（Wang Xingxing）上月表示，該公司已在機器人上測試 OpenAI 的 GPT-6，結果令人鼓舞。Unitree founder Wang Xingxing said last month the company had tested OpenAI's GPT-6 on robots with encouraging results.

他表示，如果 GPT-7 明年發布，可能為機器人產業帶來自己的「ChatGPT 時刻」—— 能力出現階躍式變化，類似 ChatGPT 在2022年推出後，讓生成式 AI 成為主流消費與商業技術。He said GPT-7, if released next year, could bring robotics its own "ChatGPT moment" — a step-change in capability akin to ChatGPT's 2022 launch, which turned generative AI into a mainstream consumer and business technology.

王興興表示，當機器人能根據語音指令，在80%的陌生環境中完成80%的任務時，他就會認為已達到這個門檻。Wang said he would consider that threshold reached when robots were able to complete 80% of tasks in 80% of unfamiliar environments from spoken instructions.

但顧問施蒂勒表示，示範中的突破不一定代表機器人已準備好廣泛商用。But consultant Stieler said a breakthrough in demonstrations would not necessarily mean robots were ready for broad commercial use.

他說：「如果問題是一個模型何時能率先在示範中處理廣泛的陌生任務，樂觀派很可能是對的。如果問題是客戶何時能放心仰賴它，我會再加上好幾年，而不是好幾個月。」"If the question is when a model first handles a broad range of unfamiliar tasks in demonstrations, the optimists may well be right," he said. "If the question is when a customer can count on it, I would add years, not months."

（蘿莉・陳與愛德華多・巴普蒂斯塔報導；凱特・梅伯里編輯）(Reporting by Laurie Chen and Eduardo Baptista; Editing by Kate Mayberry)