Embodied AI company Kinetix AI has closed an Angel+ series round of more than 500 million yuan. Investors in the round include Vertex Ventures, the venture arm of Singapore's Temasek, as well as Fortune Capital and Wanshi Capital. China Renaissance acted as exclusive financial adviser.

The proceeds will go mainly toward accelerating the company's "data-model-hardware" technology cycle and moving its embodied AI products into engineering and commercial deployment.

Founded in 2025, Kinetix AI positions itself as a full-stack embodied AI foundation model company. Its team comes out of Huawei's autonomous driving unit, the University of Hong Kong and the humanoid robotics industry, with experience spanning AI algorithms, robot hardware, data collection and product engineering.

The company's lineup so far includes a native embodied foundation model, a multimodal first-person data collection system, an embodied AI infrastructure platform, the KAI Bot humanoid robot and the KAI Hand dexterous hand.

"Data-Model-Hardware" as a Technology Flywheel

Kinetix AI argues that embodied AI cannot advance on the strength of a single model or a single robot body alone. What is needed, the company says, is a continuous feedback loop linking data, models and hardware.

Data determines a model's training quality and how well it generalizes. The model interprets the environment, plans tasks and controls the robot. The hardware carries out those tasks in the real world and feeds both successes and failures back into training.

That gives companies a closed loop: human data collection, processing on AI infrastructure, model learning and generalization, robot task execution, and real-world feedback flowing back in. The aim is to convert the physical skills humans have accumulated into capabilities a robot can learn and perform.

KAI Halo Lite Captures First-Person Data

On the data side, Kinetix AI has launched KAI Halo Lite, a first-person data capture device.

According to the company, KAI Halo Lite records human visual and motion data in real environments, then applies spatial localization, human pose estimation and 3D reconstruction algorithms to supply training data for full-body motion modeling, environment modeling and embodied AI algorithms.

Once the device is worn, users can view the captured feed on a companion terminal, along with a dynamic 3D reconstruction generated after spatial localization and pose recognition.

First-person data more faithfully captures the viewing angle, movement path and environmental interaction of a person performing a task, Kinetix AI says, which should make robots more efficient at learning from human demonstrations.

KAI World Model Ties Together Perception, Planning and Control

At the model layer, Kinetix AI has introduced KAI World Model.

The company says the model combines video generation, 4D spatial modeling and embodied task planning, supporting locomotion, manipulation and combined locomotion-manipulation tasks under a single representation.

Unlike models trained for one specific robot or a fixed setting, KAI World Model is trained across tasks and across robot bodies, an approach intended to improve how well it transfers between different robot platforms and task environments.

The company also showed the system opening bottles with a robotic arm and playing high-speed table tennis. The table tennis system, called SMASH, brings together high-speed visual perception, real-time trajectory prediction, whole-body motion control and embodied decision-making, coordinating joints across the shoulder, elbow, wrist, waist and legs.

Kinetix AI says SMASH has been deployed and tested on multiple robot bodies, and demonstrated Unitree's G1 and AgiBot's Expedition A3 executing the same shot-making routine on identical algorithms.

These results come from company demonstrations and internal testing, however, and the specific performance figures still await independent testing and validation in real deployments.

KAI Bot Built for Close Human Resemblance

On the hardware side, Kinetix AI has launched the KAI Bot humanoid robot.

The company says KAI Bot stands about 173 cm tall, weighs roughly 70 kg and has 117 degrees of freedom, with body proportions and joint placement close to a human's. The high degree-of-freedom design is meant to reduce losses when transferring human motion data to the robot.

KAI Bot can walk, grasp objects and interact with people. In public demonstrations, it picked up objects on voice command and handed them to audience members.

The company also launched KAI Hand, a dexterous hand aimed at finer object manipulation. Kinetix AI showed it grasping, gripping and absorbing high-force impacts, though performance specifications and long-term reliability still need standardized testing to confirm.

Deliveries to Industrial Customers Underway

Kinetix AI says KAI Hand and KAI Halo Lite are now on sale globally, with some industrial customer deliveries completed.

Data capture devices and dexterous hands slot into industrial customers' R&D, training and testing workflows more easily than a full humanoid robot, and they give the company a channel for gathering real-world data.

Kinetix AI's humanoid robot algorithms have been shown at the World Artificial Intelligence Conference, the World Robot Conference and the World Humanoid Robot Games, in use cases spanning exhibitions, tourism and entertainment interaction, and robot competition.

At the 2026 World Humanoid Robot Games, a joint HKU-Kinetix team fielded robots running SMASH in a table tennis exhibition. The company says the robots moved laterally, scooped up low balls and executed chop shots, showing a solid grasp of dynamic interaction.

A Full-Stack Bet That Still Needs Commercial Proof

Kinetix AI's approach spans data collection, model training, robot hardware and deployment, with the goal of using full-stack coordination to shorten the path from R&D to a shipping embodied AI product.

Embodied AI remains early, though. High degree-of-freedom hardware adds pressure on control, thermal management, reliability and cost, and a foundation model's lab performance does not necessarily translate into stable work in messy real-world settings.

Several things still need to be demonstrated in the next phase.

First, whether the company can keep sourcing high-quality real-world data. Second, whether its models transfer reliably across different robot bodies. Third, whether the robots can run over long stretches in industrial, commercial and household settings. Fourth, whether the loop between data, models and hardware can become a repeatable commercial delivery model.

With the new funding, Kinetix AI plans to increase spending on its native embodied foundation model and embodied AI infrastructure, push the joint iteration of data, models and robot hardware, and expand real-world validation of its products.