Matrix Supermind Launches MATRIX-3 Robot, Eyes L2 Autonomy by Year-End

On May 18, Matrix Supermind, an embodied AI robotics company, unveiled its next-generation humanoid robot MATRIX-3 at the 2026 Tech Day (AI DAY) event. CEO Zhang Haixing told澎湃新闻 and other media outlets that embodied AI is transitioning from L1 to L2 level this year, with the company targeting L3-level autonomy (requiring minimal human intervention in most cases) by around 2028.

Product Specifications and Pricing

The MATRIX-3 stands 170 cm tall and weighs 65 kg, with a 4-hour battery runtime and 20-minute fast-charge capability. The robot integrates Matrix Supermind’s proprietary WAVE physics foundation model and Nvidia AGX compute platform, featuring custom-designed bionic linear joints. The core drive sections—such as the thighs—employ a parallel three-linear-joint design combined with multiple sets of planetary roller screws, delivering peak thrust of 5,000 N.

The MATRIX-3 is equipped with a 27-degree-of-freedom dexterous hand (MATRIX HAND) and 3D bionic fabric “human-like musculature.” Pricing begins at ¥580,000, with the PRO version starting at ¥680,000. Both models include a one-year basic service package.

Company Background and Team

Matrix Supermind was founded in 2024 with team members recruited from Tesla, Huawei, Nvidia, and OpenAI. Zhang Haixing previously led Tesla’s China Design Center.

Cost Reduction and Market Timeline

Zhang attributed the current high price primarily to joint components, noting that individual screw costs exceed the price of an Apple computer. He stated that while China recently entered this field and remains in a learning iteration phase, the industry’s cost reduction speed is rapid—achieving 30–40% annual cost reductions despite limited production volumes. He compared the current high-price phase to early periods of smartphones, automobiles, and computers, predicting significant price declines within 3–5 years.

Zhang revealed that Matrix Supermind plans to introduce a family companion robot within 2–3 years at “very affordable” pricing, designed to be genuinely useful to consumers.

Zhang characterized 2026 as a critical mass-production year for embodied AI but emphasized the industry remains in an extremely early stage, comparable to the automobile industry over 100 years ago. He noted that leading domestic EV brands produce 2–3 million vehicles annually, yet no robotics company has reached that scale—most have not even achieved 100,000 units—and production consists of limited SKUs (stock-keeping units). This, he argued, indicates robotics is not yet a rigid consumer need and lacks a clear product-market fit or explosive growth trigger, representing lower-quality growth overall.

Matrix Supermind showcased its full-chain autonomous manufacturing system and mass-production closure at the event. Leveraging its MFH factory in Shanghai’s Zhangjiang district, MATRIX-3 has entered industrial-scale delivery. Initial capacity is planned at 10,000 units annually, with production scaling to a new magnitude within 18 months.

Business Model Evolution

Regarding business strategy, Zhang predicted that while hardware currently dominates revenue for most robotics companies, long-term hardware will represent less than 50% of revenue. The future model, he stated, will shift to “trading tokens for labor,” with operational costs, inference fees, and electricity becoming primary revenue components.

Zhang disclosed that AI and algorithm R&D investment currently exceeds 60% of the company’s budget, with hardware investment proportionally lower. “AI investment is unlimited in cost,” he stated, emphasizing his belief that robots will restructure all industries.

Zhang advised enterprises not to delay adoption due to current costs: “Companies should first get on board and start using [robots], regardless of current pricing, because prices will inevitably decline yearly. If you wait until prices become very cheap to act, you’ll have failed to integrate embodied AI with your business systems, and you’ll be eliminated. Some are already far ahead—by then they could deploy 100 or 1,000 robots and completely disrupt an entire industry.”

Autonomy Levels and Remote Control Debate

As robots deploy more widely, public skepticism about “remote control” persists. Recently, overseas embodied AI company Figure livestreamed a robot sorting parcels and challenged human efficiency, with the robot exhibiting human-like gestures such as scratching its head—reigniting “teleoperation” concerns.

Zhang explained that current robots operate in parallel manual and automatic dual modes: manual mode requires human assistance, while automatic mode enables fully autonomous operation. The industry is currently transitioning from L1 to L2, making dual-mode coexistence a normal interim state. Current products have reached L1.5 level, with formal L2 upgrade expected by year-end.

Zhang assessed Figure’s livestream as validating that “data accumulation” drives vertical-scenario automation, but noted that models optimized deeply for specific scenarios over six months show relatively weak generalization. Regarding the human-like gesture controversy, he stated: “When robots exhibit such actions, it stems from training data derived from human behavior—it’s a normal learning outcome.”

Data Collection and Industry Dynamics

Data collection remains an industry focal point. Zhang characterized it as a phase-dependent challenge, arguing that world models will eventually generate more reliable synthetic data, with human input reserved for high-precision calibration.

With major tech companies, automakers, and hundreds of startups now entering the embodied AI space, Zhang acknowledged competition is omnipresent and multidirectional. He emphasized that enterprises face resource constraints and should prioritize perfecting core products, establishing delivery channels, and achieving scalable growth before expanding SKUs. On international expansion, Matrix Supermind remains cautious: launching overseas before stabilizing domestic delivery would overextend operations and dilute resources, reducing overall efficiency.

Regarding long-term industry structure, Zhang predicted that within five years, North American companies may retain a three-month lead in model development, but China already holds a commanding advantage in robotics hardware and manufacturing.