⚛️量子位•Stalecollected in 69m
Chinese Embodied AI Startup Secures Multi-Million Funding
💡A major funding event in China's embodied AI sector focusing on the critical 'human learning' training paradigm.
⚡ 30-Second TL;DR
What Changed
Secured several hundred million RMB in new funding.
Why It Matters
This funding signals growing investor confidence in the 'human learning' approach for robotics, potentially shifting the industry focus from pure simulation to more human-aligned training data.
What To Do Next
Monitor the company's research publications for new datasets or training methodologies related to human-in-the-loop robot learning.
Who should care:Founders & Product Leaders
Key Points
- •Secured several hundred million RMB in new funding.
- •Pioneering the 'human learning' paradigm for embodied AI in China.
- •Focusing on human-centric perspectives to improve robot interaction and task execution.
🧠 Deep Insight
Web-grounded analysis with 12 cited sources.
🔑 Enhanced Key Takeaways
- •The startup, identified as 深度机智 (DeepMatrix), has cumulatively secured hundreds of millions of RMB in funding from over ten prominent investment institutions, including state-owned platforms, leading financial firms, and industrial capital, with existing shareholders continuing to invest.
- •DeepMatrix has developed the PhysBrain 1.0 embodied general intelligence foundation model system, which operates on an 'understand first, then execute' principle, and has achieved top rankings or leading positions in five major international benchmarks: WorldArena, SimplerEnv, RoboTwin 2.0, RoboCasa, and LIBERO.
- •The company's 'human learning' approach leverages first-person human perspective data to provide real-world human experience, enabling its PhysBrain 1.0 model to achieve higher learning efficiency, outperforming models trained with tens of thousands of hours of real machine data using only thousands of hours of human data.
- •DeepMatrix's models have demonstrated emergent 'flexible adaptation' capabilities, and the company is building a full-stack closed-loop system encompassing data, models, algorithms, and systems, with anthropomorphic robot designs serving as verification platforms for model migration to real robots.
- •The broader Chinese embodied AI sector experienced a significant funding surge in early 2026, raising over CNY 34.5 billion (USD 5 billion) in the first 100 days, with a notable shift in investment towards 'robot brains' and high-quality training data providers rather than solely hardware manufacturers.
📊 Competitor Analysis▸ Show
| Company | Core Approach/Focus | Key Features/Benchmarks | Funding (Latest/Total) |
|---|---|---|---|
| 深度机智 (DeepMatrix) | 'Human Learning' paradigm; 'Understand first, then execute' | PhysBrain 1.0 embodied general intelligence foundation model; Top rankings in WorldArena, SimplerEnv, RoboTwin 2.0, RoboCasa, LIBERO; Thousands of hours of human data outperform tens of thousands of hours of real machine data. | Hundreds of millions RMB (cumulative) |
| 灵初 (Lingchu) | Human operation data-driven; Real human behavior modeling | Utilizes 100,000 hours of high-quality human operation data; Self-developed exoskeleton gloves for sub-millimeter precision; Optimized WAM model architecture for <100ms reaction time; Psi-R2 (learning) and Psi-W0 (assistance) models. | Not specified in search results for latest round, but focuses on data infrastructure |
| 它石智航 (Tashi Zhihang) | 'Human-centric' data collection paradigm; 'Robot brain' pre-training | Claims 'global first embodied large model that can work'; Guinness record in technology; Focus on pre-training for robot brains. | $455 million (Pre-A round, April 2026) |
| UBTech Robotics | Humanoid robots; Multimodal reasoning models | Walker S2 (first humanoid robot capable of autonomously changing batteries); Deployed in Zeekr factory; Strategic partnership with Honda for industrial manufacturing and warehouse logistics. | Listed on Hong Kong Stock Exchange (Dec 2025); High salaries for AI talent |
| 地瓜机器人 (D-Robotics) | Robot software/hardware universal base provider; 'Chip+algorithm+software+development tools' | Focuses on 'brain' and development platform; Core products include旭日 (Xuri) series intelligent computing chips, RDK series robot developer kits. | $270 million (cumulative B-round, April 2026) |
🛠️ Technical Deep Dive
- PhysBrain 1.0 Embodied General Intelligence Foundation Model System: DeepMatrix's core technological offering, designed to enable robots to first understand the physical world and then execute tasks.
- 'Understand First, Then Execute' Paradigm: A foundational principle guiding the development of their models, contrasting with traditional methods that often rely on extensive imitation learning without deep comprehension.
- First-Person Human Perspective Data: The primary data acquisition method, providing robots with real-world human experiences to build a comprehensive understanding of physical laws and task requirements.
- Model Architecture for Physical World Understanding: Converts human experiences into capabilities for understanding, predicting, and generalizing tasks within the physical world.
- Anthropomorphic Robot Design: Utilizes robot bodies with human-like designs to facilitate the seamless transfer and verification of model capabilities from simulation or human data to real-world robot execution.
- Full-Stack Closed-Loop Development: DeepMatrix emphasizes an integrated approach covering data collection, model development, algorithm design, and system implementation from its inception.
- Superior Learning Efficiency: Achieves high performance with significantly less data; specifically, thousands of hours of human data yield results comparable to or better than tens of thousands of hours of real machine data.
- Emergent 'Flexible Adaptation' Capabilities: The developed models exhibit the ability to adapt flexibly to unforeseen circumstances, moving beyond rigid, pre-programmed task execution.
- Core Model Matrix: DeepMatrix has developed a suite of core models based on the 'human learning' route, demonstrating strong performance across Vision-Language Models (VLM), Vision-Language-Action (VLA) models, world models, and spatial intelligence.
🔮 Future ImplicationsAI analysis grounded in cited sources
The 'human learning' paradigm will become a dominant approach in embodied AI development, particularly in China.
DeepMatrix's significant funding and benchmark achievements, coupled with similar approaches from other highly funded Chinese companies, suggest a strong validation and increasing adoption of human-centric data and learning methods for embodied intelligence.
Investment in embodied AI will increasingly shift from hardware to 'robot brains' and high-quality data infrastructure.
The recent surge in Chinese embodied AI funding highlights that the industry's bottleneck has moved from hardware capabilities to the ability of AI systems to learn and reason, making data and advanced models the new focus for investors.
Embodied AI, driven by 'human learning' and robust data, will accelerate the deployment of intelligent robots in complex industrial and service scenarios.
The focus on understanding human experience and achieving flexible adaptation through 'human learning' aims to enable robots to perform complex, generalized tasks in real-world environments, addressing practical industry needs beyond simple repetitive actions.
⏳ Timeline
2025-05
DeepMatrix established (approximate, based on 'established just one year ago' in May 2026 reports).
2025-12
DeepMatrix released its first multimodal data based on human first-person perspective.
2026-03
DeepMatrix released its PhysBrain 1.0 embodied general intelligence foundation model system.
2026-04-12
Competitor Lingchu (灵初) announced using 100,000 hours of human operation data, showcasing a similar data-driven, human-centric approach.
2026-04-16
Competitor 它石智航 (Tashi Zhihang) announced a record-breaking $455 million Pre-A funding round, also claiming a 'human-centric' data collection paradigm.
2026-05-15
DeepMatrix announced cumulative funding of hundreds of millions of RMB, attracting over ten investment institutions.
📎 Sources (12)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 量子位 ↗