AI Wiki

Explore AI Wiki

Citation and evidence

Beijing Institute for General Artificial Intelligence

16 min full readUpdated 31 references

This article's verification

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

AI ResearchChinese AIEmbodied AIResearch Organizations

Cite this article

The Beijing Institute for General Artificial Intelligence (BIGAI; Chinese: 北京通用人工智能研究院, often shortened to 通研院) is a non-profit research institute in Beijing that works toward artificial general intelligence (AGI). It was set up in 2020 with backing from the Beijing municipal government, the Ministry of Science and Technology and the Ministry of Education, and it is directed by Zhu Songchun (Song-Chun Zhu), a computer vision researcher who returned to China from the United States that year.[1][2] BIGAI works closely with Peking University and Tsinghua University, and its stated aim is to build "general intelligent agents" with human-like core cognitive abilities.[1]

BIGAI's public flagship is Tong Tong (通通), a simulated child-like agent first shown in January 2024. Its published research covers 3D scene understanding, embodied AI and humanoid control, simulation platforms, AGI evaluation, automated geometry reasoning and AI agents.[3][4] The Georgetown University Center for Security and Emerging Technology (CSET) described it in 2023 as China's state-backed program aimed openly at AGI.[2]

Overview

Native name北京通用人工智能研究院
AbbreviationBIGAI (通研院)
TypeNon-profit "new-type" research and development institution
Established2020 (operations began October 2020, per CSET)
DirectorZhu Songchun
Executive deputy directorDong Le
LocationFuture Science and Technology Building, 2 Yiheyuan Road, Haidian District, Beijing
Main academic partnersPeking University, Tsinghua University
Websitebigai.ai

Expanded article table

BIGAI's Chinese "About us" page calls it a non-profit, "world-class new-type research and development institution" (新型研发机构) set up under the guidance of the Beijing Municipal Party Committee and government and promoted by the Beijing Municipal Science and Technology Commission. It lists support from Beijing, the Ministry of Science and Technology and the Ministry of Education, and partnerships with Peking University and Tsinghua University.[1] Its English motto is "Pursuing A Unified Theory of Artificial Intelligence to Create General Intelligent Agents for Lifting Humanity."[1]

The institute presents its approach as a deliberate alternative to scaling large language models. BIGAI describes a "small data, big task" paradigm and says it wants to explore the "dark matter" of intelligence.[1] CSET's 2023 report noted that BIGAI was by choice not pursuing very large language models and was instead working on "small data, big task" research grounded in brain cognition and neuroscience.[2] In a July 2026 interview with 36Kr, Zhu described the institute's "CUV" framework, which models an agent by its cognitive architecture (C), its capabilities (U) and its value system (V).[5]

History

Founding (2020)

According to CSET, Chinese newspapers announced the institute's establishment in September 2020, and it began operating in October 2020 under the sponsorship of the Ministry of Science and Technology, the Ministry of Education and the Beijing municipal government.[2] Its site in Haidian District sits next to Peking University and Tsinghua University, which CSET said makes it easier for the three institutions to work together.[2] In a director's message on the BIGAI website, Zhu writes that he returned to China full-time in September 2020 after 28 years of study and work in the United States, and then set up BIGAI with support from Beijing, the Ministry of Science and Technology and the Ministry of Education.[1]

CSET's report, written by Huey-Meei Chang and William C. Hannas and published in May 2023, said the institute had recruited some 30 top scientists educated at leading U.S. and UK universities. It also reported that Zhu had compared the importance of the AGI project, in remarks to high-level state bodies, to China's "two bombs, one satellite" programs.[2] The report pointed to claims that the facility would eventually hold 1,000 researchers.[2]

Tong Tong and the Zhongguancun Forum (2024-2026)

BIGAI unveiled Tong Tong at the Frontiers of General Artificial Intelligence Technology Exhibition in Beijing on 28-29 January 2024. The South China Morning Post reported that visitors could interact with the virtual girl: when set to like tidiness, she straightened a crooked picture frame on her own, and she found a towel to clean up spilled milk. Unlike popular large language models, the report said, she could assign tasks to herself.[6] At the 2024 Zhongguancun Forum annual conference, Tong Tong was named one of the forum's top 10 significant scientific and technological achievements. The Beijing government's announcement called it "the world's first embodied general-purpose artificial intelligence system prototype" and said that, on BIGAI's rating standard and TongTest platform, it behaved like a 3-4-year-old child.[7] China Daily reported that, according to BIGAI, Tong Tong's daily training used no more than 10 A100 chips.[8]

BIGAI has presented Tong Tong at the Zhongguancun Forum each year since 2024, releasing new versions in 2025 and 2026:

DateEventAnnouncement
28-29 January 2024Frontiers of General AI Technology Exhibition, BeijingTong Tong first shown publicly[6]
April 20242024 Zhongguancun Forum (AI Day, 27 April)Tong Tong named one of the forum's top 10 achievements[7]
29 March 20252025 Zhongguancun Forum, AI Theme DayTong Tong 2.0 released[9][10]
29 March 20262026 Zhongguancun Forum, General Artificial Intelligence ForumTong Tong 3.0, the "Tong Brain" (通脑) embodied-intelligence engine and TongAgents 2.0 released; an industry alliance launched[11][12]

Expanded article table

At the 2025 launch, Chen Hao, executive director of BIGAI's Advanced Technology Center, said Tong Tong 2.0 could plan tasks in parallel and kept its language consistent with its embodied actions, and that BIGAI planned to connect its "brain" to a physical robot.[9] Science and Technology Daily reported that Tong Tong 3.0, released on 29 March 2026, has improved cognitive, social and spatial intelligence according to Zhu. BIGAI researcher Fan Lifeng said it can read and watch videos found through search to acquire knowledge, and the report put its mental level at that of a 5-6-year-old child.[11] These age comparisons come from BIGAI's own evaluation framework and have not been independently replicated.

Organization and people

Zhu Songchun is BIGAI's director. BIGAI's website also lists him as a chair professor at Peking University and a chair professor of basic sciences at Tsinghua University.[1] In 2026 coverage he was also identified as dean of Peking University's School of Intelligence Science and Technology.[11] Dong Le serves as executive deputy director.[8][12] Siyuan Huang directs BIGAI's Embodied Robotics Center (具身机器人中心) and is a corresponding author on its UniFP and OmniXtreme papers.[13][15]

CSET's 2023 report profiled several researchers who joined BIGAI, several of them alumni of the University of California, Los Angeles, where Zhu had led a research lab. Among them were Fan Lifeng, Huang Siyuan, Liang Yitao, Liu Hangxin, Liu Tengyu, Peng Yujia, Zheng Zilong and Zhu Yixin, along with Peking University faculty such as Yang Yaodong and Zhang Muhan.[2]

Research directions and laboratories

BIGAI's research department page lists seven research directions, two technical "missions" and three laboratories:[3]

CategoryUnits
Research directionsGeneral vision; robotics; natural language; cognitive computing and commonsense reasoning; machine learning; multi-agent systems; simulation and interaction
Technical missionsTongAgents (行业智能体工场, an industry agent "factory"); TongRobot (通智大脑, a general embodied robot "brain")
National key laboratoryNational Key Laboratory of Cross-Media General Artificial Intelligence (跨媒体通用人工智能全国重点实验室), approved by the Ministry of Science and Technology and co-built with Peking University
Beijing key laboratoriesBeijing Key Laboratory of AI Safety and Standards (with Beijing University of Posts and Telecommunications, based at Peking University); Beijing Key Laboratory of Embodied Intelligent Systems (with Tsinghua University's Department of Automation)

Expanded article table

BIGAI's papers usually give the affiliation "State Key Laboratory of General Artificial Intelligence, BIGAI".[13][14] CSET noted that it was the only state key laboratory to carry "AGI" in its name.[2]

Education programs

CSET described an undergraduate "AGI experimental class" (通用人工智能实验班, nicknamed the "Tong class") drawn from Peking University's Yuanpei College and Tsinghua University's Department of Automation, taught partly by BIGAI scientists.[2] BIGAI also runs the "Tong Program" (通计划), a joint doctoral training program that BIGAI says the Ministry of Education approved and that it leads together with Chinese universities.[4] Several authors of BIGAI's 2025-2026 robotics papers are Tong Program doctoral students.[13][15]

Research areas

Embodied agents and 3D scene understanding

BIGAI researchers have published several 3D vision-language models and datasets. LEO, presented at ICML 2024, is an "embodied multi-modal generalist agent" that is trained in two stages: 3D vision-language alignment, then 3D vision-language-action instruction tuning. It covers 3D captioning, question answering, embodied reasoning, navigation and manipulation.[16] SceneVerse (ECCV 2024) introduced what its authors called the first million-scale 3D vision-language dataset: about 68,000 indoor scenes and 2.5 million vision-language pairs, which it used for a pre-training framework called GPS (Grounded Pre-training for Scenes).[17] Its successor, SceneVerse++ (CVPR 2026), uses an automated pipeline to reconstruct unlabeled internet videos into 3D scenes. BIGAI reports that it produced 6,687 real indoor scenes, 632,000 spatial question-answer samples and 9,631 navigation trajectories.[18]

Robotics and humanoid control

In 2024 BIGAI and Unitree Robotics set up a joint laboratory for embodied AI and humanoid robots, and papers from both organizations list it as an affiliation.[15][13] UniFP, a single reinforcement learning policy that learns force and position control for legged robots without force sensors, won a Best Paper Award at the Conference on Robot Learning (CoRL) 2025.[13][19] The authors report that adding its force estimates to imitation learning raised success rates by about 39.5% on four contact-rich tasks compared with position-only policies.[13]

OmniXtreme (February 2026), written with co-authors from Unitree Robotics including Wang Xingxing, trains a flow-matching policy on many motions and then refines it with actuation-aware reinforcement learning. The paper reports extreme motions such as flips on a Unitree G1.[15] In October 2026 BIGAI researchers, with Beijing Institute of Technology collaborators and Siyuan Huang as a corresponding author, released Zero-HLM, a real-to-sim-to-real framework that builds training scenes from smartphone captures. It reports zero-shot transfer of whole-body loco-manipulation policies to a Unitree G1 using only simulated demonstrations.[20]

Simulation platforms

TongSIM, released as open source in December 2025, is a general-purpose platform for training and evaluating embodied agents. Its paper describes more than 100 multi-room indoor scenes and an open-ended outdoor town; BIGAI's announcement gives the indoor count as 115 and says the platform is built on Unreal Engine.[21] BIGAI uses TongSIM as the environment for the university track of its 2026 National General Artificial Intelligence Challenge.[22]

AGI evaluation

The "Tong test," published in the journal Engineering (volume 34, 2024) by Peng Yujia, Zhu Songchun and colleagues, proposes evaluating AGI through "dynamic embodied physical and social interactions" (DEPSI) instead of a conversational Turing test. It sets out five critical characteristics as AGI benchmarks and a value- and ability-oriented system with five levels of AGI milestones in a virtual environment that can generate unlimited tasks.[23]

Reasoning and mathematics

TongGeometry, by BIGAI and Peking University researchers, appeared as the cover article of Nature Machine Intelligence in January 2026. The paper describes a neuro-symbolic system that uses guided tree search to both propose and prove olympiad geometry theorems. The authors report that it solved all 30 problems in the IMO-AG-30 benchmark, and that three of its proposed problems were used in a Chinese national team qualifying exam and a U.S. olympiad.[24] As reported by the South China Morning Post, the authors also said the benchmark run took 38 minutes on a machine with 32 CPU cores and one RTX 4090 GPU. They added that they do not claim TongGeometry beats an average IMO gold medallist at geometry in general. The same report quoted the authors as saying Google DeepMind's AlphaGeometry "successfully solved only three of our 10 proposals".[25]

BIGAI researchers also co-authored Absolute Zero, a self-play reasoning method developed with Tsinghua University and Penn State researchers; BIGAI's Zilong Zheng was one of its two corresponding authors.[26]

AI agents (TongAgents)

TongAgents is BIGAI's agent platform. BIGAI describes it as a hierarchical multi-agent system that separates a planner, executors and an independent verifier.[27] In March 2026 BIGAI said TongAgents ranked first on AssistantBench and on tau2-bench across the retail, airline and telecom domains, second on Terminal-Bench 2.0 and in the top three on Mind2Web 2. These are BIGAI's own claims.[27] As of 30 October 2026, the public Terminal-Bench 2.0 leaderboard listed TongAgents with Gemini 3.1 Pro at 80.2% (plus or minus 2.6), tied for sixth place.[28]

BIGAI researchers, with Humanlaya, xbench and others, also built $OneMillion-Bench, a set of 400 expert-curated tasks in law, finance, industry, healthcare and natural science. It scores agents with rubrics and measures task value in dollars of expert time; the paper lists NeurIPS 2026 (Evaluations and Datasets Track).[29]

Notable projects

ProjectYearTypeNotes
Tong test2023-2024AGI evaluation frameworkEngineering perspective article[23]
LEO2023 (ICML 2024)3D embodied generalist agentTwo-stage 3D VL alignment and VLA tuning[16]
SceneVerse2024 (ECCV 2024)3D vision-language datasetAbout 68K scenes, 2.5M pairs[17]
Tong Tong2024-2026Simulated child-like general agentVersions 2.0 (2025) and 3.0 (2026)[6][9][11]
UniFP2025Legged force-position control policyCoRL 2025 Best Paper Award[19]
TongSIM2025Embodied-agent simulation platformOpen source on GitHub and Hugging Face[21]
TongGeometry2026Olympiad geometry proposer and solverNature Machine Intelligence cover article[24]
OmniXtreme2026Humanoid high-dynamic motion controlWith Unitree Robotics[15]
TongAgents2026Multi-agent system and agent platformVersion 2.0 released March 2026[12][27]
$OneMillion-Bench2026Expert-task agent benchmark400 tasks across five domains[29]
SceneVerse++2026 (CVPR 2026)3D data from internet videos6,687 reconstructed scenes[18]
Zero-HLM2026Real-to-sim-to-real humanoid loco-manipulationUnitree G1 experiments[20]

Expanded article table

Collaborations and spin-offs

  • Universities. BIGAI shares researchers and students with Peking University and Tsinghua University, and it co-builds its national key laboratory with Peking University and one of its Beijing key laboratories with Tsinghua's Department of Automation.[2][3]
  • Unitree Robotics. The two set up a joint laboratory in 2024 and co-author papers including UniFP and OmniXtreme.[15][13]
  • Leju Robotics. For the 15th National Games torch relay in Shenzhen on 2 November 2025, Leju Robotics named China Mobile, Harbin Institute of Technology and BIGAI as ecosystem partners that gave technical support for its Kuavo humanoid, whose running motion-control algorithm had been optimized.[30] BIGAI says Unitree and Leju were among the members of the TongRobot (通智大脑) alliance it formed in 2025.[13][15]
  • Delta Intelligence. Delta Intelligence (北京德塔源创智能科技有限公司, 德塔智能) is a humanoid foundation-model company that was formally established in January 2026 and was incubated by BIGAI, according to an April 2026 Zhongguancun report.[31] Beijing News also listed it among the embodied-AI companies BIGAI has incubated.[12]

Continue reading

More to read

Topics:AI ResearchChinese AIEmbodied AIResearch Organizations

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7Beijing Institute for General Artificial Intelligence, "关于我们" (About us), accessed 30 October 2026. bigai.ai/about
  2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Huey-Meei Chang and William C. Hannas, "Spotlight on Beijing Institute for General Artificial Intelligence," Center for Security and Emerging Technology, May 2023. cset.georgetown.edu/...ral-artificial-intelligence (doi:10.51593/20230013)
  3. ^1 ^2 ^3Beijing Institute for General Artificial Intelligence, "科研部门" (Research departments), accessed 30 October 2026. bigai.ai/%e7%a7%91%e7%a0%94%e9%83%a8%e9%97%a8
  4. ^1 ^2Beijing Institute for General Artificial Intelligence, home page, accessed 30 October 2026. bigai.ai
  5. ^"Zhu Songchun: China's AI Industry Should Not Be Misled by the 'Musk Belief'" (interview), 36Kr, 21 July 2026. eu.36kr.com/...3904621311756161
  6. ^1 ^2 ^3Zhang Tong, "World's first AI child unveiled in Beijing by award-winning Chinese scientist who left US for China," South China Morning Post, 2 February 2024. scmp.com/...-creates-worlds-first-ai-child-beijing
  7. ^1 ^2"World's First General-purpose Intelligent Digital Person 'Tongtong' Selected as One of Top 10 Significant Scientific and Technological Achievements at 2024 ZGC Forum Annual Conference," Beijing Municipal Government, 9 May 2024. english.beijing.gov.cn/...t20240509_3667994
  8. ^1 ^2Yuan Shenggao, "Impact of AI felt throughout five-day event," China Daily, 6 May 2024. mobile.chinadaily.com.cn/...37f11ced50dde094231c67
  9. ^1 ^2 ^3"China unveils upgraded version of world's first AGI agent," Bastille Post, 30 March 2025. bastillepost.com/...sion-of-worlds-first-agi-agent
  10. ^Liu Zhen and Song Chen, "记者手记：遇见通用人工智能体" (Reporter's notes: meeting a general AI agent), Xinhua, 30 March 2025. news.cn/...c
  11. ^1 ^2 ^3 ^4"全球首个通用智能人'通通'3.0亮相" (Tong Tong 3.0 debuts), Science and Technology Daily via Beijing International Science and Technology Innovation Center, 30 March 2026. ncsti.gov.cn/...t20260330_242361
  12. ^1 ^2 ^3 ^4Wei Yingzi, "通研院董乐：人工智能的原始创新不能纯靠打价格战" (BIGAI's Dong Le on original AI innovation), Beijing News, 30 March 2026. bjnews.com.cn/...1774841764129482
  13. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Peiyuan Zhi, Peiyang Li, Jianqin Yin, Baoxiong Jia and Siyuan Huang, "Learning a Unified Policy for Position and Force Control in Legged Loco-Manipulation," arXiv:2505.20829, 2025; and BIGAI news release on the CoRL award, 9 October 2025. arxiv.org/...2505.20829
  14. ^Zhe Sun et al., "TongSIM: A General Platform for Simulating Intelligent Machines," arXiv:2512.20206, December 2025. arxiv.org/...2512.20206
  15. ^1 ^2 ^3 ^4 ^5 ^6 ^7Yunshen Wang et al., "OmniXtreme: Breaking the Generality Barrier in High-Dynamic Humanoid Control," arXiv:2602.23843, February 2026; and BIGAI news release, 5 March 2026. arxiv.org/...2602.23843
  16. ^1 ^2Jiangyong Huang et al., "An Embodied Generalist Agent in 3D World," arXiv:2311.12871 (ICML 2024). arxiv.org/...2311.12871
  17. ^1 ^2Baoxiong Jia et al., "SceneVerse: Scaling 3D Vision-Language Learning for Grounded Scene Understanding," arXiv:2401.09340 (ECCV 2024). arxiv.org/...2401.09340
  18. ^1 ^2Yixin Chen et al., "Lifting Unlabeled Internet-level Data for 3D Scene Understanding," arXiv:2604.01907 (CVPR 2026); and BIGAI news release on SceneVerse++, 30 April 2026. arxiv.org/...2604.01907
  19. ^1 ^2"Awards," CoRL 2025 (Conference on Robot Learning, Seoul). 2025.corl.org/...awards
  20. ^1 ^2Dianyi Yang et al., "Zero-HLM: Generalizable Humanoid Loco-Manipulation without Real Robot Data," arXiv:2610.35075, 26 October 2026. arxiv.org/...2610.35075
  21. ^1 ^2"通研院AI'全能物理空间训练场'TongSIM开源上线" (TongSIM released as open source), BIGAI news, 29 December 2025. bigai.ai/...8a%e7%ba%bf%ef%bc%8c%e4%b8%a4%e5%a4%a9 ; code github.com/...tongsim
  22. ^"2026全国通用人工智能挑战赛正式启动" (2026 National General AI Challenge launched), BIGAI news, 10 July 2026. bigai.ai/...bd%e6%8c%91%e6%88%98%e8%b5%9b%e6%ad%a3
  23. ^1 ^2Yujia Peng et al., "The Tong Test: Evaluating Artificial General Intelligence Through Dynamic Embodied Physical and Social Interactions," Engineering 34(3): 12-22, 2024. doi:10.1016/j.eng.2023.07.006. engineering.org.cn/...j.eng.2023.07.006
  24. ^1 ^2Chi Zhang, Jiajun Song, Siyu Li, Yitao Liang, Yuxi Ma, Wei Wang, Yixin Zhu and Song-Chun Zhu, "Proposing and solving olympiad geometry with guided tree search," Nature Machine Intelligence, January 2026 (publication page). yzhu.io/...iq2026nmi ; preprint arXiv:2412.10673
  25. ^"Chinese AI goes next level in geometry at a top US maths Olympiad," South China Morning Post, republished by Peking University News, 29 January 2026. newsen.pku.edu.cn/...15392
  26. ^Andrew Zhao et al., "Absolute Zero: Reinforced Self-play Reasoning with Zero Data," arXiv:2505.03335, May 2025. arxiv.org/...2505.03335
  27. ^1 ^2 ^3"产品能力领跑全球！通研院 TongAgents 登上多项国际智能体榜单" (TongAgents tops several international agent leaderboards), BIGAI news, 20 March 2026. bigai.ai/...b9%e5%9b%bd%e9%99%85%e6%99%ba%e8%83%bd
  28. ^"Terminal-Bench 2.0 leaderboard," tbench.ai, accessed 30 October 2026. tbench.ai/...2.0
  29. ^1 ^2Yang Liu et al., "$OneMillion-Bench: How Far are Language Agents from Human Experts?," arXiv:2603.07980, March 2026. arxiv.org/...2603.07980
  30. ^"Humanoid 'Kuafu' debuts as special torchbearer at 15th National Games relay in Shenzhen," Global Times, 2 November 2025. globaltimes.cn/...1347146.shtml
  31. ^"北京通研院孵化！德塔智能连续三轮融资超亿元" (Delta Intelligence, incubated by BIGAI, completes three funding rounds), Zhongguancun via Beijing International Science and Technology Innovation Center, 16 April 2026. ncsti.gov.cn/...t20260416_244072

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

1 revision · v1 · 3,128 words · full history

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

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: History, Tong Tong releases, research projects and agent claims checked against BIGAI's own pages, CSET, SCMP, Xinhua, China Daily, Beijing government releases, the cited arXiv papers and the Terminal-Bench 2.0 leaderboard, October 2026.

Cite this page: AI Wiki. "Beijing Institute for General Artificial Intelligence." aiwiki.ai, updated 29 Oct 2026, fact-checked 29 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/bigai

Edit article