PRESS RELEASE


Technology Frontier Co., Ltd. | 日本語

HUMANOID CONTROL AI / VISION-LANGUAGE-ACTION

PATENT PENDING

Humanoid Control AI “TF-VLA”
Seeking hardware makers for collaboration and a joint venture

Ask in plain words, and the humanoid walks over, finds the object, picks it up and carries it. We have developed the control AI that serves as its “brain,” and filed a patent application for it. The next step is a real robot. We are looking for companies that can design and develop the body (hardware) of a humanoid to build this business with us, from joint validation on real hardware to the establishment of a joint venture.

SUMMARY

Summary

Technology Frontier Co., Ltd. (“the Company”) has developed TF-VLA, a control AI for humanoid robots that hierarchically integrates a Vision-Language-Action (VLA) model, which generates robot motion from camera images and human language, with physics-based walking control. The Company has filed a patent application for its core technology (patent pending).

Given an instruction such as “Put the red block in the left tray” or “Climb the stairs,” in Japanese or English, TF-VLA has the humanoid walk to the work table, find the target, grasp it and carry it, all autonomously. In physics simulation we have confirmed a 93% success rate on the carrying task, walking over a course that includes stairs and a slope, and getting up after a fall.

What we have is the robot’s “brain.” Bringing it to society requires a “body” that actually moves. Starting today, we are therefore seeking partners that can design and develop the body (hardware) of a humanoid, for joint validation on real robots with a view to establishing a joint venture for commercialization.

BACKGROUND

Background: a humanoid can only work once the “brain” and the “body” come together

With labor shortages continuing in manufacturing, logistics, nursing care and agriculture, expectations are rising for humanoids that can work in the same environments as people. Getting there is hard in different ways for the brain (control AI) and for the body (hardware).

Teaching every motion one by one does not scale

Conventional robots need specialist engineers to program each motion. Work on real sites is varied and keeps changing. Being able to simply ask in words is a precondition for wide adoption.

Running AI output as it is would be unsafe

The output of a learned model is probabilistic and sometimes includes motions that cannot be executed. To work beside people, safety has to be ensured by a mechanism outside the AI.

Walking, carrying and falling must be handled as one

Task AI and walking control are often researched separately. An architecture that connects the two and keeps the task going, including recovery from a fall, has yet to be established.

A brain alone or a body alone is not a product

Control AI has to be trained to fit the characteristics of the hardware, and hardware only shows its value once it has control AI. Working together from an early stage is the short path to practical use.

FEATURES

Key features of TF-VLA

TF-VLA brings instruction understanding, action generation, safety checking, walking and recovery from a fall together in a single control system.

01

Works from spoken instructions

It understands instructions in Japanese and English and splits them into manipulation and walking tasks. It handles sequences such as “first … then …” and asks back when an instruction is ambiguous.

02

Walking and carrying as one

A layered design: the VLA model handles work that needs vision, and physics-based control handles walking. Approaching the table, crossing stairs and slopes, and returning to the start are executed as one flow.

03

Safety is not left to the AI

AI output is treated as a proposal. Before execution, a safety guard checks speed, workspace, self-collision and the approach of people. The emergency stop cuts power without passing through the AI.

04

Recovers instead of stopping

Even when part of the AI’s output has a problem, the system has ways to keep the overall task going. When a situation is truly dangerous it stops reliably and does not resume on its own.

05

Gets back up after a fall

If pushed over, the robot gets up by itself, returns to walking and resumes the task. Time spent on the ground is not counted against the task’s time limit.

06

Can be trained for your hardware

Walking policies are acquired by reinforcement learning in physics simulation. We provide a pipeline for retraining with a different robot model, along with a simulator and an operator panel for validation.

INTELLECTUAL PROPERTY

The core technology of TF-VLA is patent pending

PATENT PENDING

We have filed a patent application for technology obtained through the development of TF-VLA. For any collaboration, we will make the handling of intellectual property clear up front so that partners can take part in joint development with confidence.

Control system and control method

An invention concerning an architecture that hierarchically integrates everything from spoken instructions through action generation, safety checking and execution to walking control.

Safety together with continuity

An invention concerning processing that keeps a task going while checking the output of a learned model, and an architecture that builds recovery from a fall into the task.

Policy training method

An invention concerning methods for training, in physics simulation, policies for walking over terrain such as stairs and for getting up after a fall.

* Details of the application are explained after a non-disclosure agreement is signed. “Patent pending” does not guarantee that a patent will be granted.

RESULTS

Results (measured in physics simulation)

We have confirmed the following in physics simulation using a publicly available model of a commercial humanoid. Validation on real hardware will be carried out with the partners we are seeking here.

93%

Carrying task success

From a spoken instruction to walking over, grasping and placing. 56 of 60 trials

0.18 s

AI decision time

Average time to decide the next motion. Design target: within 0.5 s

16 / 16

Getting up after a fall

Times the robot stood up by itself after being pushed over in 8 directions

0.22 m

Stopping distance

Maximum distance from a stop command while walking to standstill. Requirement: within 0.3 m

Other validated items
WalkingRound trips over a course with 10 cm stair risers and a 15-degree slope; walking backward, sideways and turning; holding its footing when pushed
Instruction understandingManipulation and walking instructions in Japanese and English, splitting of sequential instructions, asking back for missing information
SafetyChecks before execution, monitoring during execution, slowing and stopping when a person approaches, emergency stop (power cut that bypasses the AI), no automatic resumption
Quality106 automated tests, all run on every change

* Figures are the Company’s own measurements as of September 2026, obtained in physics simulation. They do not guarantee performance on real hardware. Part of the validation uses data restricted to research use, which will be replaced with commercially usable data before commercialization.

PARTNERSHIP

Call for collaboration and joint venture partners

We want to combine our “brain” with a partner’s “body” and jointly develop a humanoid that can work on real sites. Once joint validation shows promise, we would like to discuss establishing a joint venture as the vehicle for commercialization.

WHAT WE BRING

The brain (control AI)

  • The complete TF-VLA control AI and its patent-pending technology
  • Training of walking policies fitted to your robot model
  • A physics-simulation validation environment and operator panel
  • Safety design (pre-execution checks, runtime monitoring, stop sequences)

WHAT WE LOOK FOR IN A PARTNER

The body (hardware)

  • Design, development and prototyping of humanoid hardware
  • Actuators, reducers, sensors, electrical systems, batteries
  • A test environment for real hardware and manufacturing know-how with volume production in view
  • Expertise in mechanical and electrical safety of the robot

Who we are looking for

  • Companies that can design and develop the body (hardware) of a humanoid
  • Robot makers, and also companies in automotive, industrial machinery, precision equipment and similar fields with strength in mechanism and drive design
  • Companies developing key components such as actuators, reducers and sensors
  • Companies willing to consider joint validation on real hardware and the commercialization beyond it

Forms of collaboration

  • Co-development of hardware and control AI under a joint development agreement
  • Joint validation on real hardware (proof of concept) and sharing of the results
  • Establishment of a joint venture for commercialization (ownership ratio and roles to be agreed)
  • Handling of intellectual property is set out in a contract before the collaboration begins

How the collaboration proceeds

  1. Initial discussion

    We sign a non-disclosure agreement and confirm each other’s technology and goals.

  2. Fit check in simulation

    We bring your robot model into our environment and check its fit for walking and manipulation.

  3. Joint validation on hardware

    Under a joint development agreement, we validate walking, manipulation and safety on the real robot step by step.

  4. Commercialization and joint venture

    Based on the results, we discuss the form of commercialization, including establishing a joint venture.

* The terms of any collaboration or joint venture are decided through individual discussion. This release is not a commitment to conclude a contract on any particular terms.

ROADMAP

Next steps

Together with our partners, we plan to work on the following (timing to be agreed).

Validation on real hardware

Step-by-step validation of walking, manipulation, safety stops and recovery from a fall on the partner’s robot.

A wider range of tasks

From carrying blocks to real on-site work. We will broaden the objects and tasks to match the partner’s intended applications.

Field trials

We aim for trials at sites short of hands, such as manufacturing, logistics, nursing care and agriculture.

FAQ

Frequently asked questions

What does the pending patent cover?

We have filed an application covering a control system and control method for humanoid robots, and a method for training policies such as walking. Details of the application are explained after a non-disclosure agreement is signed.

Has TF-VLA run on a real robot?

Not yet. All validation so far has been done in physics simulation. We intend to carry out validation on real hardware together with the partners we are seeking here.

What kind of companies are you looking for?

Companies that can design and develop the body (hardware) of a humanoid. This is not limited to robot makers; it includes companies with design and development capability in structures, actuators, reducers, sensors and electrical systems.

Can it be adapted to our own robot?

Yes. If you provide a model of your robot (for example in URDF or MJCF format), we can start by training a walking policy and checking the fit in our simulation environment.

Are the terms or ownership ratio of the joint venture already decided?

No. We would like to begin with joint validation and then agree on the form and terms of commercialization based on its results.

TECHNOLOGY

Technology overview

NameTF-VLA (humanoid control AI)
Intellectual propertyPatent pending (control system, control method, policy training method)
ComponentsInstruction understanding, action generation by a VLA model, safety guard, motion execution, physics-based walking control, recovery from a fall, monitoring and logging, operator panel
Supported instructionsJapanese and English (carrying objects, moving, crossing stairs and slopes, sequential instructions)
Validation environmentPhysics simulation (humanoid model with 29 degrees of freedom)
Development stageValidation in simulation is complete; now seeking partners for validation on real hardware
What we are seekingCollaboration with companies that can design and develop the body (hardware) of a humanoid, and the establishment of a joint venture

CONTACT

Inquiries about collaboration and a joint venture

Tell us about your hardware, your technology and the applications you have in mind. We will propose a demonstration in simulation and a way to proceed. Press inquiries are also welcome at the address below.

info@technologyfrontier.co.jp
Company profile
CompanyTechnology Frontier Co., Ltd.
RepresentativeKazuya Sato, Representative Director and CEO
AddressHulic Minatomirai 10F, 1-1-7 Sakuragicho, Naka-ku, Yokohama, Kanagawa 231-0062, Japan
Websitehttps://www.technologyfrontier.co.jp/
Contactinfo@technologyfrontier.co.jp

* Company and product names mentioned are trademarks or registered trademarks of their respective owners. The results in this release were measured in physics simulation and do not guarantee performance on real hardware. “Patent pending” does not guarantee that a patent will be granted. This page is an English translation of the Japanese release; the Japanese text prevails in case of any discrepancy.