Manifold AI

Company Overview

Manifold AI (Beijing Manifold Space Technology Co., Ltd.), founded in Beijing on May 22, 2025, is a company focused on embodied intelligenceworld modelA cutting-edge AI company focused on R&D and implementation. The company has a clear positioning—it is China’s first startup to independently develop a world model and directly deploy it as an embodied foundation model in robotics applications. Rather than remaining confined to academic papers and datasets, it is committed to ”engineering and commercialization.”

Dr. Wu Wei, Founder and CEO, joined SenseTime’s founding team in 2015. He led the research, development, and commercialization of SenseTime’s Kaiwu World Model and guided his team to victory in the Waymo SimAgents Challenge for two consecutive years, defeating teams such as Tesla. A co-founder is a Tenured Professor at Tsinghua University (a Changjiang Scholar appointed by the Ministry of Education) who led his laboratory in establishing WorldArena, an evaluation platform for embodied world models. Team members hail from top institutions such as Tsinghua University, Peking University, Shanghai Jiao Tong University, and UCLA. They have held key technical positions at leading companies including Microsoft, SenseTime, Momenta, and Xpeng, and have collectively published over 200 papers in top-tier conferences, which have been cited more than 100,000 times.

As of June 18, 2026, just one year after its founding, the company has completed six rounds of financing, with cumulative funding in the Pre-A round reaching nearly 1 billion yuan. Its valuation has surpassed the unicorn threshold, and its investors include top-tier institutions and industrial capital firms such as Legend Capital, Huawei Habor, Huakong Fund, China Reform Fund, Yifeng Capital (a Temasek subsidiary), and BAIC Industrial Investment.

Development History: Six Rounds in One Year—An Astonishing Pace

Going from zero to unicorn status within a year, the company’s funding pace was closely aligned with its product development pace. Wu Wei himself has stated, ”The company aims to be a product-driven company.”

Products and Services: A World Model Matrix Covering Outdoor, Indoor, and Airspace Environments

Manifold AI is committed toWorld Model Action (WMA) Technical Roadmap, building on its self-developed world model, the company has established a comprehensive product portfolio for embodied intelligence:

this isThe industry’s first company to offer an embodied world model that simultaneously covers outdoor, indoor, and airspace environments.

Real-world scenariosFrom the lab to the marketplace:

• E-commerce Logistics: Warehouse robot cargo recognition, path planning, and material handling and sorting
• 3C Manufacturing: Precision Operations and Quality Inspection
• Automotive Manufacturing: Training in Robot Operation Skills for Production Lines
• MR Pan Entertainment: Exploring Mixed Reality Applications

The company has been the first to connect toNVIDIA Jetson Thor(NVIDIA’s developer kit designed specifically for physical AI), for the edge-based deployment of embodied world models.

Core Technology: Not a ”video generator that produces visually appealing content,” but an ”interactive physics engine”

1. WorldScape: The world’s first real-time interactive world model

In stark contrast to the approach of traditional video generation models—which prioritize ”rendering over rules”—WorldScape places greater emphasis onUnified Modeling of Physical Laws, Spatial Structures, and Multi-Source Interactions::

• Unified Action—A Framework for Modeling the State of the World: Incorporate spatial displacement and object interaction into a single generation process to avoid inconsistencies caused by splicing multiple modules together
• Spatial Representations and Constraints in 3D Geometric Perception: Explicitly introduce geometric constraints to ensure that the generated results maintain a consistent spatial structure throughout the iterative process.
• Near-real-time inference speed: Interactive generation at 6–16 FPS on a single GPU
• Generating Interactive Spaces from a Single Image: The ability to generate an interactive 3D scene from a single image

2. WorldScape Policy: From ”Understanding the World” to ”Precise Control”

Built on the WorldScape foundation, WorldScape Policy uses the world model to predict spatiotemporal states and combines this with visual input to perform spatial endogenous reasoning in order to execute actions:

• Accuracy is superior to existing VLA models across the board
• PossessLow-Sample and Zero-Sample Execution Capabilities(Context-Based Learning, without the need for large amounts of ontology data)
• It demonstrates exceptional resistance to interference and robustness even in complex environments with significant random variations in object position, lighting, and background.
• Significant improvements compared to the VLA approach in long-range, high-precision missions

3. MoE Architecture: A Creative Breakthrough in World Modeling

WorldScape version 0.2 was the first to introduceMixture of Experts Architecture, This is an original contribution by Manifold AI:

pass (a bill or inspection etc)Adaptive Integration of Gating Mechanisms, while maintaining parameter scalability and preventing interference between different domains of knowledge.

4. Hardware-Data-Model Closed-Loop System

The company has independently developedEgo-centric, UMI, Hardware-in-the-Loop RL...and various other integrated hardware-software data acquisition devices to build a model-driven data pipeline, with a daily production capacity exceeding100,000 clips, has accumulatedHundreds of thousands of hoursHigh-quality data. This infrastructure serves as the foundation for the company’s ongoing technological evolution.

5. WorldArena: Redefining the Standards for World Model Evaluation

Manifold AI has partnered with dozens of research institutions in China and abroad to establishWorldArena, the International Standardized Assessment Benchmark, covering six major dimensions and 16 indicators:

WorldArena reveals a significant relationship between visual effects and embodied functions“The Perception-Function Gap”, shifting the industry's evaluation criteria from ”how realistic the generated video is” to“Is the model truly applicable to robotic decision-making?”. The first round of evaluations covered 14 representative models, including those from teams such as NVIDIA, Google, Alibaba, and Li Feifei’s World Labs. The CVPR 2026 WorldArena Challenge received a total of over 200 submissions.

Future Prospects: At the Paradigm Shift of Embodied Intelligence

1. Industry Background: The VLA approach has hit a bottleneck, and the world model has emerged as the new consensus

In the first half of 2026, the embodied intelligence sector had invested heavily for three years, but VLA models performed poorly in real-world environments. Tong Ti, a partner at Matrix Partners China, stated bluntly: ”Relying solely on data patterns and correlation learning is no longer sufficient to meet the needs of next-generation intelligent systems.” An investor from Inno Fund offered an even sharper assessment: ”Causality is a unique expression of human intelligence; causal intelligence is thousands or even tens of thousands of times more data- and parameter-efficient than statistical empirical correlations.”

Manifold AI's Roadmap—Replacing General-Purpose VLMs with World Models as Foundation Models——is becoming the new consensus in the industry.

2. Competitive Landscape: Global Top Tier

WorldScape topped both the WorldScore and WorldArena rankings—two authoritative lists—with other contenders during the same period including:

Manifold AI’s Key Differentiators:Despite having specifications that are only 10%—a fraction of its competitors’—it has remained at the top of the rankings for nearly two months....demonstrating world-leading spatial intelligence density.

3. Commercialization Strategy: ”Climb to the summit, laying eggs along the way”

The company is pursuing a two-pronged strategy:

• Short-term: Commercialization of models in areas such as RoboScape and AirScape, with implementation in scenarios such as e-commerce logistics and 3C manufacturing
• Mid-term: Expanding from Robotics to Broader Physical AI Scenarios
• Long-term: To increase the number of in-house developed and empowered robots to exceed 10% of the total market volume, becoming Physical AI AgentInfrastructure Provider

4. Investor Confidence: Collective Entry of Industrial Capital

In this round of financing,BAIC Industrial Investment, Xinneng Venture CapitalThe entry of industrial capital is of great significance—it marks the recognition of the world-model technology approach by the real economy. The Pre-A round, which raised a total of nearly 1 billion, demonstrates that the capital markets now view Manifold AI as a key player in the embodied intelligence sector.