A robot approaches a patient's bed. It does not lift anything heavy. It stops at a respectful distance, turns a screen toward the person, and waits. In a hospital, that restraint may be more valuable than strength.
Fourier Intelligence (傅利叶), a Shanghai company founded in 2015, has chosen this narrow path. Its GR-3 humanoid (人形机器人) is designed for care: soft outer shell, touch-sensitive skin, a display where a face would be, and a body that is deliberately unthreatening. The company made its name in rehabilitation robotics before entering humanoids, and that lineage shapes everything about the product.
Key takeaways
- Fourier began in rehabilitation robots and exoskeletons in 2015, then entered humanoids with the GR-1 in 2023.
- The GR-3 is 165 cm tall and weighs about 71 kg, with roughly 3 kg of payload per arm.
- It was launched in August 2025 and made its US debut at CES in January 2026.
- A rehabilitation project in Malaysia, launched in June 2026, includes Fourier systems; the humanoid part remains at proof-of-concept stage.
- Care robots face certification and trust barriers that cannot be solved by hardware improvements alone.
A company that came from hospitals
Most humanoid makers arrived from artificial intelligence labs, automotive engineering, or consumer electronics. Fourier arrived from clinics. Its early products were lower-limb exoskeletons and rehabilitation systems for people recovering from spinal cord injury or stroke. The company says its equipment reaches more than 2,000 institutions in over 40 countries, a footprint built over a decade of selling into regulated medical environments.
That history explains the GR-3's odd specifications. A payload of about 3 kg per arm is unimpressive next to industrial machines that lift tens of kilograms. Fourier is not trying to move a patient. It is trying to be present, to interact, and to be safe enough to stand near a vulnerable person. The soft shell is not cosmetic; it is a statement about what happens if the robot falls.
The technical choices follow the same logic. According to the company's specification table, the robot uses 31 distributed pressure sensors, a four-microphone array, an eye display, and RGB plus structured-light face tracking. Its control stack reportedly uses a two-path language layer: fast rule-based responses for simple commands, and a full language model for contextual interaction. Two hot-swappable battery packs give roughly three hours of operation.
The Malaysia project as a test case
In June 2026, a rehabilitation centre operated by Malaysia's social security institution, PERKESO, opened with Fourier's biomechanical analysis and training systems among its equipment. Reporting from southern China described the project as being worth tens of millions of yuan and said the centre would also introduce Fourier humanoids for cognitive training and companion tasks.
Fourier describes the humanoid portion as still at proof-of-concept stage: clinical and local validation first, then registration and compliance certification, then sales. That sequence is telling. The company is not claiming a breakthrough in autonomy. It is describing the slow institutional work of becoming a vendor in a foreign healthcare system.
Fourier established a Singapore headquarters in 2018 and a Malaysian team of several dozen people. Its experience there reflects a general rule for medical robotics: the product is easy to demonstrate and hard to register. Training scenarios had to be customised to Kuala Lumpur street scenes so that rehabilitation exercises would transfer to patients' real environments. That is the kind of detail that never appears in a launch video but determines whether a system is useful.
Why care is a harder market than factories
Factory customers buy by throughput. A robot either improves output per shift or it does not. Care customers buy by risk reduction, staff relief, and patient engagement, all of which are slower to measure and harder to attribute.
The numbers reflect that. Fourier had raised nearly ¥800 million (≈ US$110M) in a Series E round back in January 2025, well before the GR-3 launch. In July 2026, Chinese financial media reported that the company had begun preparations for a Shanghai STAR Market listing at a valuation above ¥8 billion (≈ US$1.12B). The GR-3 is not in mass production, and no official price has been published; distributors and reports place indicative figures around US$120,000, which should not be treated as a quote.
The company also develops open platforms, including a humanoid called N1 and an action dataset, and its software stack supports Isaac Lab and MuJoCo for simulation. Those efforts matter because care robots will need to learn tasks from limited demonstration data. But a well-documented platform is not the same as a deployed clinical workflow.
Honest limitations
- The GR-3's payload, battery life, and sensor counts come from Fourier's own specification table and reviews of it. I found no independent operational trial in a working care facility.
- The Malaysia project's value was reported as "tens of millions of yuan" by one provincial outlet; I could not confirm a precise figure from a primary source.
- The reported STAR Market preparation and valuation come from Chinese financial media, not a regulatory filing. No listing has been confirmed.
- Pricing figures are indicative distributor estimates, not company prices.
- Claims about the scale of Fourier's rehabilitation business are company statements.
- I have deliberately excluded a widely circulated report about a Meta prototype, because Reuters said it could not verify it and neither company commented.
What readers can do now
- If you evaluate care technology, ask what the robot is certified to do, not what it demonstrated. Those are different lists.
- For families considering robotic companionship for an older relative, check whether the device has any clinical validation for the specific condition involved.
- Watch the regulatory pathway, not the product launch. In healthcare, approvals move slower than prototypes, and they are the real gate.
