Skip to main content

BeingBeyond (智在无界)

Category: Robotics / Embodied AI

Chinese embodied AI startup building general-purpose foundation models for humanoid robots using large-scale human video data BeingBeyond (智在无界) was founded in 2025. The company is led by 卢宗青 (Zongqing Lu). Based in Beijing, China. Team size: 11-50. Latest round: Seed. Key investors include Legend Star (联想之星), Zhipu AI Z-Fund / Star Capital (星连资本/Z基金), Yanyuan Venture Capital (燕缘创投), Binfu Capital (彬复资本), LH Ventures, Koala Fund, Linkerbot.

Founded
2025
Headquarters
Beijing, China
Team size
11-50

Value proposition

Developing implicit world models and general-purpose foundation models for humanoid robots that can run on edge devices (100 TOPS chips), dramatically reducing deployment costs to ~$20/month per robot — 98% cheaper than NVIDIA Cosmos

Products and solutions

Being-H series (hand motion generation model for dexterous manipulation), Being-M series (whole-body motion generation model), Being-VL series (multimodal foundation model for robotics), Being-W series (low-level whole-body control model), Being-H-Flash (implicit world model for edge deployment), Being-Actor (whole-body teleoperation system), Being-Dex (dexterous manipulation solution), D1 (desktop dexterous robotic arm), U1 (Real DexUMI)

Unique value

First company globally to achieve real-time implicit world model deployment on 100 TOPS-level edge chips; uses large-scale human video data (200K+ hours) for pre-training instead of expensive robot data; cross-embodiment generalization; monthly deployment cost as low as ~$20/robot

Target customer

Humanoid robot manufacturers, industrial automation companies, logistics/warehousing operators, and robotics integrators seeking embodied AI solutions

Industries served

Embodied AI / Robotics, Humanoid Robots, Industrial Automation, Logistics & Warehousing, Manufacturing

Technology advantage

Implicit world model architecture (latent space reasoning instead of pixel-level video generation); Universal Async Chunking (UAC) inference technology; 200K+ hours of first-person human video + 15K+ hours robot demonstration data; cross-embodiment generalization; supports both NVIDIA and domestic Chinese AI chips; 70%+ R&D staff, ~60% PhDs from top universities

How they differentiate

Only company deploying implicit world models on edge (100 TOPS) vs competitors' cloud-dependent explicit world models; 98% cheaper than NVIDIA Cosmos; uses internet-scale human video data (not expensive robot data); cross-embodiment generalization; supports domestic Chinese chips alongside NVIDIA

Main competitors

NVIDIA Cosmos (explicit world model), 穹彻智能 (Noematrix/Noematrix Brain), 星动纪元 (GalaxyBot/Star Dynamics), 自变量机器人 (AutoBot/X-Automation)

Key partnerships

Collaborating with leading robot manufacturers for scenario validation, Peking University (founder's academic affiliation), Beijing Academy of Artificial Intelligence (Zhiyuan Institute - core team source), Legend Star (联想之星) - lead investor, Zhipu AI Z-Fund (智谱Z基金) - investor

Notable customers

Working with leading robot manufacturers for scenario validation (specific names not publicly disclosed)

Major milestones

Jan 2025: Company founded, 2025: Released Being-H0 (1K hours human video pretraining), 2025: Released Being-H0.5 (10K hours), 2025: Released Being-M0 with MotionLib (first million-scale motion generation dataset), 2025-06: Completed seed round (Legend Star-led), 2026-01: Completed angel round (LH Ventures, Koala Fund, Linkerbot), 2026-04: Released Being-H0.7 (200K hours, top-ranked on 6 benchmarks), 2026-06: Released Being-H-Flash (first implicit world model running on 100 TOPS edge chips)

Growth metrics

Founded Jan 2025; 4 model generations in 1 year (Being-H0 → H0.5 → H0.7 → H-Flash); 20K+ hours human video → 200K+ hours training data scale-up

Market positioning

Early-stage Chinese embodied AI startup positioned as a pure-play "robot brain" / foundation model provider (not robot hardware maker), competing with both international (NVIDIA Cosmos) and domestic (Noematrix, GalaxyBot) players in the embodied world model space

Geographic focus

China (primary), with potential global expansion; competes in the global embodied AI foundation model market

About 卢宗青 (Zongqing Lu)

Tenured Associate Professor, School of Computer Science, Peking University (since 2024, joined PKU as Assistant Professor in 2017); Former Head of Multimodal Interaction Research Center, Beijing Academy of Artificial Intelligence (BAAI/Zhiyuan Institute); Postdoc at Pennsylvania State University; PhD from Nanyang Technological University (2014); Master & Bachelor from Southeast University

Official website: