AI Wiki

PsiBot

AI CompaniesChinese AIEmbodied AIRobotics Companies
13 min read
Updated
Suggest editHistoryTalk
RawGraph

Last edited

Fact-checked

Sources

24 citations

Revision

v2 · 2,573 words

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

FieldValue
CompanyPsiBot (灵初智能, Lingchu Zhineng, "Lingchu Intelligence")
Registered name北京灵初智能科技有限公司 (Beijing Lingchu Intelligence Technology Co., Ltd.)
FoundedSeptember 2024
Founder and CEOWang Qibin (王启斌)
Co-foundersChen Yuanpei (陈源培), Chai Xiaojie (柴晓杰)
HeadquartersRegistered in Beijing (Haidian); company site lists Xuhui district, Shanghai
SectorEmbodied AI, dexterous manipulation, robotics
Key modelsPsi R0, Psi R0.5, Psi R1, Psi R2, Psi W0
Key hardwarePsi V1 wheeled robot, Psi H1 dexterous hand, Psi E1 exoskeleton
Funding2 billion yuan ($280M) disclosed March 2026; ~$100M round in progress at $1.48B valuation (Bloomberg, July 2026)
Websitepsibot.ai

PsiBot (Chinese: 灵初智能, pinyin Lingchu Zhineng, rendered in English as Lingchu Intelligence) is a Chinese embodied AI startup founded in September 2024 that builds reinforcement learning based vision language action models and dexterous hand hardware for long-horizon manipulation tasks[17][18]. The company drew international attention on July 23, 2026, when Bloomberg reported it was raising close to $100 million at a $1.48 billion valuation in a round led by carmaker Chery Automobile, which would bring its total funding to about $300 million in under two years[2]. PsiBot is unusual among China's robot startups in that it began with manipulation models and dexterous hands rather than a full humanoid robot, and it has published a series of named models (Psi R0 through Psi R2) trained end to end with reinforcement learning[5][8][15].

What it is

PsiBot describes itself as a company focused on general-purpose embodied intelligence, large vision language action (VLA) models, and dexterous manipulation algorithms, built as a full-stack loop from models to self-developed hardware, simulation platforms, and data systems[17][18]. Its stated mission is to "create infinite productivity with AI and robots"[17].

The company's technical bet is that the bottleneck in embodied AI is manipulation, not locomotion. Rather than racing to ship a bipedal humanoid, PsiBot concentrated on two-handed dexterous operation: grasping, scanning, packing, tying bags, and other multi-step tasks in open environments, executed by wheeled dual-arm platforms and third-party robot bodies running its models[5][8][9]. Chinese tech media summarized the positioning at founding as building "the industry's leading general-purpose dexterous manipulation agent," deliberately avoiding core component manufacturing, mobile chassis work, and a rush to full humanoids[3][13].

Commercially the company targets business customers first, in manufacturing (incoming material inspection, handling, packaging) and logistics (sorting and distribution). In a June 2025 interview, founder Wang Qibin said machines make economic sense when priced below roughly 24 months of local labor cost, and set a sales target of several hundred million yuan by the end of 2026, with deliveries planned to overseas warehouse customers in the United States and Europe[13]. By March 2026 the company said it had completed small-scale verification in the warehouses of real logistics customers, in semi-structured settings such as clothing warehouses where robots scan barcodes and sort items[10][12].

Founding and team

PsiBot was founded in September 2024. The company's site gives September 1, 2024 as the founding date, and the registered entity, Beijing Lingchu Intelligence Technology Co., Ltd. (北京灵初智能科技有限公司), was incorporated in Beijing's Haidian district on September 11, 2024, with Wang Qibin as legal representative[1][17][19]. The company website lists a headquarters address in Shanghai's Xuhui district (the district whose state investment vehicle later led its Pre-A round), so the company effectively operates across Beijing and Shanghai[18][10]. As of mid 2025 the team numbered around 80 people, including roughly 30 interns from its joint laboratory with Peking University[13].

The founding team combines a commercial hardware veteran with young reinforcement learning researchers:

  • Wang Qibin (王启斌), founder and CEO, has close to 20 years of experience taking consumer electronics and robots from definition to global launch, spanning BlackBerry mobile phones, Sonos smart speakers in China, and robotics at hotel-robot maker Yunji Technology and JD.com's robotics business[3][12][13].
  • Chen Yuanpei (陈源培), co-founder and reinforcement learning lead, is a Guangdong-born researcher of the post-2000 generation who was a visiting scholar at Stanford University under Karen Liu and Fei-Fei Li. Chinese press credits him with early work using reinforcement learning to control two arms and two dexterous hands simultaneously on real hardware[3][12].
  • Chai Xiaojie (柴晓杰), co-founder, spent about 15 years in robotics and autonomous driving, including experience mass-producing L4 autonomous driving products[3][10].
  • Yang Yaodong (杨耀东), an assistant professor at Peking University's Institute for Artificial Intelligence with a PhD from University College London, serves as chief scientist and leads the PKU-PsiBot joint laboratory on embodied dexterous manipulation[3][11][17].

The joint lab, plus academic cooperation the company lists with Tsinghua University, Shanghai Jiao Tong University, and the Chinese University of Hong Kong (Shenzhen), functions as PsiBot's research pipeline[18].

Technology: reinforcement learning for long-horizon manipulation

PsiBot's models are built around a "fast-slow brain" hierarchical architecture: a high-level planner based on a pretrained vision language model interprets instructions and decides strategy, while a low-level controller trained heavily with reinforcement learning in simulation executes dexterous actions with closed-loop visual feedback[7][8][9]. The company argues this mix of reinforcement learning, limited imitation learning, simulated synthetic data, and sparse real-robot data avoids the very high cost of collecting large teleoperated datasets on real hardware[13].

Named releases, all verifiable in Chinese press coverage at the time:

ReleaseDateWhat it is
Psi R0December 30, 2024Described by the company as the first end-to-end embodied model trained with reinforcement learning to complete long-horizon dexterous tasks; demonstrated on a dual-hand e-commerce workflow of grasping, barcode scanning, placing, and tying a plastic bag[4][5]
Psi R0.5March 3, 2025End-to-end VLA model built with the DexGraspVLA framework; the company said about two hours of dexterous grasping data (2,094 trajectories) generalized to thousands of objects, positions, lightings, and backgrounds[6][7]
Psi R1, Psi V1, Psi H1April 27, 2025Hierarchical end-to-end VLA plus reinforcement learning model, launched together with the Psi V1 wheeled dual-arm robot (175 cm, 32 degrees of freedom) and the Psi H1 dexterous hand (16 active degrees of freedom with tactile sensing); demonstrated playing mahjong continuously for more than 30 minutes[8][9][14]
Psi R2 and Psi W0April 10, 2026Psi R2, which the company calls the first "World Action Model" pretrained on roughly 100,000 hours of human manipulation data, takes images and language as input and outputs predicted video and robot actions; Psi W0 is an action-conditioned world model used to evaluate and improve the policy and transfer human-hand dynamics to robots via reinforcement learning[15][16]
ψ-SynRobotApril 2026Self-developed integrated robot platform combining task execution with onboard data collection; the company said mass production had started, aimed at logistics, retail, and industrial scenarios[20]

The mahjong demonstration attached to Psi R1 became the company's signature. A robot shuffled, drew, flipped, and played tiles against human opponents for over 30 minutes of continuous "chain of action thought" execution, combining human-robot, robot-robot, and robot-environment interaction[8][9]. Wang later said the demo took about a month and a half to train and existed mainly to prove long-horizon operation in open environments, though he noted mahjong skills could feed a future companion robot[13].

Data is the other half of the strategy. PsiBot built the Psi-SynEngine collection system around self-developed exoskeleton gloves (SynGlove) and the Psi E1 homomorphic exoskeleton, which the company says capture human hand trajectories at sub-millimeter precision with vision, joint angle, and tactile channels[12][15][18]. In April 2026 it open-sourced the first batch, about 1,000 hours, of its human hand manipulation dataset on Hugging Face under a CC BY 4.0 license (PsiBotAI/SynData), out of a claimed reserve of 100,000 hours, and said Psi R2 fine-tuned with fewer than 100 real robot trajectories could assemble phones, pack industrial goods, and stack boxes[15][16]. The company also said Psi R2 ranked first on the MolmoSpaces embodied AI benchmark ahead of models from Physical Intelligence and NVIDIA; these are company claims from its launch materials[15][21]. At the 2026 World Robot Conference in Beijing, PsiBot exhibited a packing robot driven by natural language commands and a delivery robot with a 7-axis arm and 6-axis dexterous hand handling packages up to 3 kg[19].

Funding history

PsiBot raised money at a pace unusual even by the standards of China's embodied AI boom: roughly 2 billion yuan (about $280 million) disclosed within 18 months of founding, followed by the Chery-led round reported in July 2026[2][10][12].

RoundAnnouncedAmountInvestorsSource
Seed (announced at the time as an angel round)November 13, 2024UndisclosedLed by GL Ventures (Hillhouse's venture arm) and Lanchi Ventures; the company's March 2026 recap lists GL Ventures, Lanchi Ventures, and robot maker AgiBot as seed investors[3][11][22]
AngelDisclosed March 10, 2026 (combined with Pre-A: 2 billion yuan total)Part of 2 billion yuanCDB Capital (国开金融, China Development Bank's investment arm), Guozhong Capital, the CCTV Media Convergence Industrial Investment Fund, the strategic arm of an unnamed listed company, a fund under fiber maker YOFC, and funds including Yuansheng Venture Capital, Zhuhai Technology Industry Group, Junshan Investment, Yanyuan Venture Capital, Dami Capital, Wofu Capital, and Binfu Capital[10][11][12]
Pre-ADisclosed March 10, 2026 (combined with angel: 2 billion yuan total)Part of 2 billion yuanLed by Shanghai state-owned Xuhui Capital, with the Liangxi Sci-Tech Innovation Phase II mother fund (managed by Bohua Capital), Wuxi Venture Capital, and market funds Pufeng Capital and Timing Capital; existing shareholders increased stakes[10][11][12]
New round (in progress)Reported July 23, 2026Close to $100 million at a $1.48 billion valuationLed by Chery Automobile, with investors including Lens Technology; Bloomberg reported the financing was close to being finalized, not closed[2]

Round labels shifted between announcements: the November 2024 financing led by GL Ventures and Lanchi Ventures was called an angel round when announced, but the company's March 2026 communications describe those investors (plus AgiBot) as the seed round and apply the "angel" label to the state-capital round[3][11][22]. No valuation was disclosed with the 2 billion yuan announcement[12]. Bloomberg's July 2026 report put cumulative fundraising at about $300 million since inception in 2024[2].

What the investor list signals

The cap table tracks the phases of China's embodied AI funding cycle. The first money came from elite venture capital (Hillhouse's GL Ventures, Lanchi Ventures) and a strategic peer, AgiBot, whose co-founder Peng Zhihui's participation was widely noted in Chinese coverage of the young founding team[3][22]. The 2025 to 2026 rounds brought in "national team" capital: China Development Bank's CDB Capital, a CCTV-affiliated media fund, and Shanghai's district-level Xuhui Capital, along with local state funds from Wuxi. Chinese press framed this as state capital backing embodied intelligence as a strategic industry, and the announcement landed during the March 2026 Two Sessions political meetings[10][11][12].

The round reported in July 2026 adds industrial strategics. Chery Automobile, the lead, is a major Chinese carmaker that has its own humanoid ambitions: its incubated Aimoga Robotics unit began selling a humanoid robot to consumers at 285,800 yuan (about $41,830) in April 2026 and said it had delivered 300 humanoids and 1,000 robot dogs across more than 30 countries by the end of 2025[23]. Lens Technology, which Bloomberg describes as a sensor maker supplying Apple and Tesla, is a consumer electronics component manufacturer that gives PsiBot a route into precision manufacturing supply chains, one of its target deployment sectors[2]. Automakers and electronics suppliers investing in manipulation-focused robotics mirrors the pattern elsewhere in the humanoid robot market, where factory and warehouse work is the first paying use case[2][24].

Position in China's embodied AI boom

PsiBot reached a $1.48 billion reported valuation in under two years without shipping a bipedal humanoid, which distinguishes it from compatriots like Unitree, UBTECH, and AgiBot that sell full robots[2][24]. Bloomberg placed the deal in a wave of Chinese AI startups hitting billion-dollar valuations in 2026 as investors chase embodied AI[2]. A commentary syndicated by Benzinga in March 2026 noted that global humanoid robot shipments reached about 18,000 units in 2025, up 508 percent year over year, with Chinese manufacturers leading volume deliveries, and read PsiBot's state-backed 2 billion yuan raise as a signal of Beijing's bet on the sector[24].

Within that boom, PsiBot's differentiation is its data and manipulation stack: exoskeleton-glove human data at scale, reinforcement learning in simulation, and models that target 30-minute-plus task horizons rather than walking demos[8][13][15]. Wang has argued the industry is moving from a "hardware wave" to a "data wave," and that driving down the cost of manipulation data matters more than rushing a humanoid form factor[21]. The company said it aimed to build China's largest dexterous hand dataset in 2026, with about 100 data-collection devices deploying in the Beijing area and a 2026 collection target of one million hours[10][12].

The approach has trade-offs that company statements themselves acknowledge. PsiBot does not make core components or mobile bases, so it depends on partners for much of the robot body[13]. Its deployment record so far consists of small-scale verifications in customer warehouses rather than large fleets, and its revenue target of several hundred million yuan by the end of 2026 is a forecast, not a result[10][13]. The headline benchmark and generalization claims for its models (the 250x data-efficiency comparison against Figure AI's dataset for Psi R0.5, the MolmoSpaces first place for Psi R2) come from the company's own launch materials and have not been independently audited[7][15]. The $1.48 billion valuation likewise rests on a round that Bloomberg reported as still being finalized as of July 23, 2026[2].

See also

References

  1. 北京灵初智能科技有限公司 (company registration entry). 百度百科, accessed July 23, 2026. https://baike.baidu.com/item/%E5%8C%97%E4%BA%AC%E7%81%B5%E5%88%9D%E6%99%BA%E8%83%BD%E7%A7%91%E6%8A%80%E6%9C%89%E9%99%90%E5%85%AC%E5%8F%B8/65107614 ↩
  2. Saritha Rai and Bloomberg News, "China's Psibot Hits $1.48 Billion Valuation in Latest AI Startup Funding Round", Bloomberg, July 23, 2026. https://www.bloomberg.com/news/articles/2026-07-23/china-s-psibot-becomes-latest-ai-startup-to-hit-1-billion-value ↩
  3. "高瓴、蓝驰领投灵初智能，致力于打造业界领先通用灵巧操作智能体", 量子位 (QbitAI), November 2024. https://www.qbitai.com/2024/11/218183.html ↩
  4. "灵初智能发布首个基于强化学习的端到端具身模型 Psi R0，双灵巧手协同进行复杂操作", IT之家, December 30, 2024. https://www.ithome.com/0/821/147.htm ↩
  5. "灵初智能发布首个基于强化学习的端到端具身模型 Psi R0", 腾讯新闻, December 30, 2024. https://news.qq.com/rain/a/20241230A05KBU00 ↩
  6. "灵初智能发布端到端 VLA 模型 Psi R0.5，仅需两小时数据实现物品、场景全面泛化", 极客公园 (GeekPark), March 3, 2025. https://www.geekpark.net/news/346512 ↩
  7. "灵初智能端到端VLA模型Psi R0.5，两小时数据实现物品、场景全面泛化", 腾讯新闻, March 3, 2025. https://news.qq.com/rain/a/20250303A06TTT00 ↩
  8. "灵初智能发布分层端到端VLA+强化学习算法模型Psi-R1", 腾讯新闻, April 28, 2025. https://news.qq.com/rain/a/20250428A052KS00 ↩
  9. "灵初智能推出端到端 VLA 模型 Psi R1，让机器人变身麻将牌友、外卖小哥以及商店理货员", 搜狐, April 2025. https://m.sohu.com/a/890606785_122178203/ ↩
  10. "灵初智能已完成20亿融资，获'国家队'资本重磅投资，助力中国具身智能高质量发展", 腾讯新闻 (华兴交易 release), March 10, 2026. https://news.qq.com/rain/a/20260310A042QK00 ↩
  11. "灵初智能完成20亿融资：蓝驰创投与智元机器人是种子轮投资", 新浪科技, March 10, 2026. http://finance.sina.com.cn/tech/csj/2026-03-10/doc-inhqnrqv6329051.shtml ↩
  12. "这个广东00后，已经融资20亿", 36氪, March 2026. https://www.36kr.com/p/3716918109730432 ↩
  13. "对话灵初智能CEO王启斌：让机器人进工厂有意义，让机器人学会打麻将也有意义", 新浪财经, June 11, 2025. https://finance.sina.com.cn/roll/2025-06-11/doc-inezupah3838669.shtml ↩
  14. Psi V1, 轮式双臂长程专家. AI星踪岛 robot directory, accessed July 2026. https://aixzd.com/robot/psi-v1 ↩
  15. "灵初智能双模型亮相，全球最大人类手部数据集正式开源", 新浪科技, April 13, 2026. https://finance.sina.com.cn/tech/roll/2026-04-13/doc-inhuipia5711457.shtml ↩
  16. PsiBotAI/SynData dataset (CC BY 4.0). Hugging Face, accessed July 23, 2026. https://huggingface.co/datasets/PsiBotAI/SynData ↩
  17. 迈向通用具身智能新纪元 (About Us). PsiBot, accessed July 23, 2026. https://www.psibot.ai/en/about-us/ ↩
  18. PsiBot home page (products: Psi R series, Psi V1, Psi H1, Psi E1, Psi Sim, Psi Data, SynGlove). https://psibot.ai/en/home/ ↩
  19. 北京灵初智能科技有限公司, exhibitor profile, 2026 World Robot Conference. https://www.worldrobotconference.com/expo/company/524.html ↩
  20. "灵初智能发布自研整机ψ-SynRobot并启动量产，正式开启具身智能规模化落地新征程", PsiBot, April 2026. https://www.psibot.ai/%E7%81%B5%E5%88%9D%E6%99%BA%E8%83%BD%E5%8F%91%E5%B8%83%E8%87%AA%E7%A0%94%E6%95%B4%E6%9C%BA%CF%88-synrobot%E5%B9%B6%E5%90%AF%E5%8A%A8%E9%87%8F%E4%BA%A7%EF%BC%8C%E6%AD%A3%E5%BC%8F%E5%BC%80%E5%90%AF/ ↩
  21. "对话灵初智能CEO 王启斌：具身智能，正从'硬件浪潮'进入'数据浪潮'", 钛媒体 (TMTPost), May 2026. https://www.tmtpost.com/7996985.html ↩
  22. "具身智能创企灵初智能宣布获高瓴、蓝驰领投天使轮融资", 新浪科技, November 13, 2024. https://finance.sina.com.cn/tech/roll/2024-11-13/doc-incvwmtt1552285.shtml ↩
  23. Phate Zhang, "Chery brings humanoid robot to general consumer market", CnEVPost, April 13, 2026. https://cnevpost.com/2026/04/13/chery-brings-humanoid-robot-to-general-consumer-market/ ↩
  24. "PsiBot's $280M Fundraising Signals China's Bet On Embodied AI", Benzinga (syndicated via Sahm Capital), March 17, 2026. https://www.sahmcapital.com/news/content/psibots-280m-fundraising-signals-chinas-bet-on-embodied-ai-2026-03-17 ↩

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

1 revision by 1 contributors · full history

Suggest edit

What links here

Humanoid robot market