CtrlK
| MATRIX-3 | |
|---|---|
 | |
| General information | |
| Manufacturer | Matrix Robotics |
| Country of origin | China |
| Year unveiled | January 2026 |
| Status | Prototype (pilot deployments mid-2026) |
| Price | ~$85,000 USD |
| CEO | Zhang Haixing (Allen Zhang) |
| Headquarters | Shanghai, China |
| Website | matrixrobotics.ai |
MATRIX-3 is a third-generation general-purpose humanoid robot developed by Matrix Robotics (formally Shanghai Matrix Superintelligent System Integration Co., Ltd.), a robotics and artificial intelligence company headquartered in Shanghai, China. Unveiled in January 2026, MATRIX-3 represents a ground-up redesign of the company's humanoid platform, incorporating advances in biomimetic sensing, dexterous manipulation, natural locomotion, and embodied AI reasoning. The robot's most notable feature is a pair of 27-degree-of-freedom (DOF) cable-driven dexterous hands, called the "Lingxi Hand," which match the approximate degrees of freedom found in a human hand and represent one of the highest hand-DOF counts among commercially targeted humanoid robots as of early 2026.
Built on a proprietary neural network architecture developed by Matrix Super Intelligence, MATRIX-3 is designed to achieve zero-shot neural generalization, meaning it can understand basic physical principles and execute novel tasks from natural language instructions without requiring task-specific training data. Matrix Robotics has positioned MATRIX-3 as a platform for transitioning humanoid robots from scripted industrial machines into truly autonomous, general-purpose physical intelligence agents.
Matrix Robotics was founded by Zhang Haixing, who also goes by Allen Zhang in English-language communications. Zhang served as the founding leader of the Tesla China Design and Research Center, where he contributed to the development of the Tesla Optimus humanoid robot program, the autonomous robotaxi initiative, electric vehicles, and smart charging infrastructure. He brings over 20 years of experience in product research and development, design, and technology entrepreneurship spanning consumer electronics, electric vehicles, and AI.
Before his tenure at Tesla, Zhang founded Crazybaby, a consumer electronics company that gained recognition for its innovative magnetic levitation Bluetooth speakers. The company's Mars speaker, which featured a flying saucer-shaped disc that levitated above a subwoofer column using patented magnetic technology, raised over $750,000 on Indiegogo in 2014. Crazybaby went on to win multiple design awards, including the German Red Dot Award for Product Design (2017), the Japan Good Design Award, and the CES Innovation Award for Best of Innovation. Zhang holds over 100 patents and has received multiple international design accolades.
Applying lessons from both Tesla's robotics program and his consumer electronics background, Zhang founded Matrix Robotics with the stated mission of building humanoids capable of handling heavy, dangerous, and repetitive tasks in unstructured, real-world environments. The company's team includes specialists in AI, humanoid robotics, and autonomous driving.
Matrix Robotics also has a partnership with Star Dynamics Technology Co., Ltd., a joint venture established in September 2025 by Hongrun Construction Group (60% stake) and Matrix Super Intelligence (40% stake). Star Dynamics focuses on the research, development, production, and commercialization of humanoid robots, with plans for mass production of its Star-1 humanoid beginning in 2026.
The Matrix-1 was Matrix Robotics' first-generation humanoid robot, debuted in mid-2025. Standing 180 cm tall and weighing 67 kg, the Matrix-1 featured 22 DOF per hand, a 2.28 kWh custom battery enabling 5 hours of operation, walking speeds up to 7.6 km/h (1.2 m/s), and 10 kg payload capacity per arm. It was equipped with 8 onboard RGB cameras, LiDAR, and ultrasonic sensors for perception, along with an 8-core processor with integrated GPU running a real-time operating system. The Matrix-1 introduced the company's embodied manipulation model (MLM2.0) and vision language model integration for autonomous operation.
MATRIX-3 represents a systematic reconstruction of both hardware and software compared to its predecessor. Key differences include:
| Feature | Matrix-1 | MATRIX-3 |
|---|---|---|
| Height | 180 cm | 170 cm |
| Total DOF | Not disclosed | 37 |
| Hand DOF (per hand) | 22 | 27 |
| Hand actuation | Force-position hybrid | Cable-driven (tendon) |
| Biomimetic skin | No | Yes (3D woven fabric) |
| Tactile sensing | Fingertip only | Full-body distributed + fingertip |
| AI generalization | Task-specific training | Zero-shot generalization |
| Walking speed | 7.6 km/h | Not disclosed |
| Battery | 2.28 kWh, 5 hr runtime | Not disclosed |
| Payload per arm | 10 kg | Not disclosed |
| Status | Prototype/demo | Prototype (pilot mid-2026) |
MATRIX-3 stands approximately 170 cm tall, with a total of 37 degrees of freedom across its body. The robot's exterior features a streamlined, futuristic design intended for safe operation alongside humans in shared spaces. The compact body eliminates the bulky rear battery enclosure of the Matrix-1, distributing components more evenly throughout the frame.
The centerpiece of MATRIX-3's hardware is the Lingxi Hand, a five-fingered dexterous hand with 27 degrees of freedom per hand. The joint architecture closely mirrors human hand anatomy, and the hand uses advanced cable-driven (tendon-driven) actuation to achieve an extremely lightweight profile while maintaining high-speed precision.
Cable-driven hand designs route actuators to the forearm and transmit force through tendons to the finger joints, similar to how human hands work. This approach offers several advantages over motor-in-joint designs: it reduces the weight and bulk of the hand itself, allows for more degrees of freedom in a compact space, and enables more natural finger movements. The Lingxi Hand's 27 DOF matches the approximate degrees of freedom of a human hand, which is generally considered to have 27 independent DOFs across the fingers and thumb.
The hand is designed to perform a wide range of tasks, including:
Each fingertip integrates ultra-sensitive tactile sensor arrays capable of detecting pressure changes as low as 0.1 newtons. This sensitivity enables the robot to perceive micro-force variations during contact, making it possible to interact safely with humans and handle fragile objects without damage.
The 27-DOF Lingxi Hand places MATRIX-3 at or near the top of commercially targeted humanoid robots in terms of hand dexterity. For reference, the human hand has approximately 27 independent degrees of freedom. As of early 2026, the landscape of humanoid hand dexterity includes:
| Robot | Manufacturer | Hand DOF (per hand) | Actuation Type |
|---|---|---|---|
| MATRIX-3 | Matrix Robotics | 27 | Cable-driven (tendon) |
| Tesla Optimus Gen 3 | Tesla | 22 | Electric motor |
| 1X NEO | 1X Technologies | 20 | Electric motor |
| Figure 02 | Figure AI | 16 | Electric motor |
| Shadow Dexterous Hand | Shadow Robot Company | 24 | Pneumatic/electric |
| Robonaut 2 | NASA/GM | 22 | Tendon-driven |
| Human hand | N/A | ~27 | Tendon (biological) |
The Lingxi Hand's 27 DOF is notable because it achieves near-human-level articulation in a commercially targeted humanoid platform, rather than in a standalone research hand. Most production humanoid robots sacrifice hand DOF to reduce cost and complexity, typically offering between 12 and 22 DOF per hand.
The MATRIX-3's locomotion and upper-body movement are powered by a proprietary set of integrated linear actuators that Matrix Robotics calls the "Bionic Super Joint" system. These actuators combine high power density with low-noise operation, providing a stable and agile foundation for full-body coordination. The linear actuator design differs from the rotary motors used in many competing humanoids, offering advantages in certain types of linear motion and force application.
The joint system supports the robot's humanoid gait, which was trained on extensive human motion-capture datasets. According to Matrix Robotics, the resulting walking and navigation behavior is intended to be indistinguishable from human movement, with natural stride patterns, balance adjustments, and turning behavior.
MATRIX-3 introduces a pioneering approach to robot surface design through its three-dimensional woven biomimetic skin. This flexible fabric covers the robot's chassis and serves multiple functions:
This is the first time biomimetic skin textures and high-dimensional tactile sensing have been deeply integrated in a Matrix Robotics product. The combination of full-body distributed sensing with fingertip precision (0.1 N sensitivity) creates a layered perception system that provides what the company describes as "physical interaction intuition comparable to human touch."
MATRIX-3 uses a multimodal perception system that combines visual and tactile inputs:
The cognitive core of MATRIX-3 runs on a proprietary neural network architecture developed by Matrix Super Intelligence. This architecture links perception, planning, and task execution into a unified system that the company refers to as its "embodied reasoning" framework. Rather than treating perception, decision-making, and motor control as separate modules, the system integrates these functions to enable more fluid and adaptive behavior.
The most significant AI capability claimed for MATRIX-3 is zero-shot neural generalization. In the context of robotics, zero-shot generalization refers to a system's ability to perform tasks it has never been specifically trained on, by applying an understanding of general physical principles and transferring knowledge from related tasks.
Traditional industrial robots and many current humanoid platforms require specific programming or training data for each task they perform. If a robot is trained to pick up a cup, it may fail when presented with a differently shaped container. MATRIX-3's neural architecture is designed to overcome this limitation by:
This capability, if fully realized, would represent a significant step toward general-purpose robot learning, where a single platform can handle diverse, unpredictable tasks in unstructured environments like homes, hospitals, and retail spaces.
MATRIX-3 accepts natural language instructions, bridging the gap between human intent and robot action. Rather than requiring specialized programming or teach-pendant operation, users can communicate tasks conversationally. The system interprets these instructions through its neural network and generates corresponding motion plans. This approach builds on broader trends in large language model integration with robotic systems, where language understanding serves as an interface layer between human operators and physical agents.
Matrix Robotics has identified several sectors for MATRIX-3 deployment:
The combination of soft biomimetic skin, sensitive touch, and zero-shot task generalization is specifically intended to make MATRIX-3 suitable for environments where close human contact is unavoidable, a domain where traditional hard-shelled industrial robots with rigid programming are poorly suited.
Matrix Robotics launched an Early Access Program for MATRIX-3 in early 2026, inviting select industry partners to participate in evaluation and pilot deployment. Initial pilot deployments are scheduled to begin in mid-2026. Zhang Haixing has stated that the company intends to control the per-unit price of its robots at a level affordable to the general public within two to three years, with the goal of bringing humanoid robots to consumer end-users.
Large-scale manufacturing of MATRIX-3 is expected to begin in the second half of 2026, though the timeline is contingent on the results of the early access program and further technical validation.
The January 2026 unveiling of MATRIX-3 attracted both excitement and skepticism within the robotics community. The announcement relied heavily on computer-generated imagery (CGI) rather than live demonstrations of physical hardware. RoboHorizon, a robotics-focused publication, noted that the presentation "remains just that, a vision, rendered in glorious CGI until a physical prototype proves otherwise," adding that the publication would "keep enthusiasm on standby until we see actual hardware moving with the grace promised in the animation."
This skepticism reflects a broader pattern in the humanoid robotics industry, where promotional materials and rendered videos frequently outpace demonstrated physical capabilities. Several Chinese and international humanoid robotics companies have faced similar criticism for marketing robots through cinematic CGI before showing real-world performance. The gap between CGI promises and hardware reality remains one of the key challenges for evaluating claims made by humanoid robot developers.
CEO Allen Zhang has acknowledged the importance of physical validation, stating: "The philosophy of MATRIX-3 is to integrate machine intelligence into human physical spaces as naturally and safely as possible." The mid-2026 pilot deployments will serve as the first opportunity for independent evaluation of MATRIX-3's real-world capabilities.
MATRIX-3 enters an increasingly crowded humanoid robotics market. As of 2026, major competitors include:
MATRIX-3 differentiates itself primarily through hand dexterity (27 DOF, the highest among these competitors), biomimetic tactile skin, and its claimed zero-shot generalization. However, the robot's lack of publicly demonstrated hardware performance and reliance on CGI marketing at launch place it behind competitors like Tesla and Figure AI that have shown extensive real-world video demonstrations.
The Chinese humanoid robotics ecosystem has grown rapidly, with local governments committing significant funding to humanoid robotics development. Matrix Robotics operates within this ecosystem alongside companies such as Unitree Robotics, Agibot, Galbot, and others racing to deliver commercially viable humanoid platforms.
| Category | Specification | Value |
|---|---|---|
| Physical | Height | 170 cm |
| Physical | Form factor | Compact (no rear protrusion) |
| Mobility | Total degrees of freedom | 37 |
| Manipulation | DOF per hand | 27 |
| Manipulation | Hand name | Lingxi Hand |
| Manipulation | Fingers per hand | 5 |
| Manipulation | Hand actuation | Cable-driven (tendon) |
| Sensing | Fingertip tactile sensitivity | 0.1 N |
| Sensing | Body tactile sensing | Distributed (biomimetic skin) |
| Sensing | Vision | Spatial perception foundation model |
| Actuators | Type | Integrated linear (Bionic Super Joint) |
| Computing | AI architecture | Proprietary neural network |
| Computing | LLM integration | Yes |
| Computing | Generalization | Zero-shot |
| Availability | Price | ~$85,000 USD |
| Availability | Status | Prototype |
| Availability | Pilot deployment | Mid-2026 |
| Availability | Target markets | Manufacturing, logistics, healthcare, service, home |