Physical AI

What we keep being asked

Robot and humanoid builders need perception inside a battery budget, without a cloud round trip. Once you write down the latency and the milliwatts, the compute question stops being a software question and becomes a silicon one.

Fig. 2The intelligence sits at the machine; the cloud stays in the background.

What it is

Physical AI is AI that acts through a body: humanoids, mobile robots, and automated machines. Every one of them is an embedded computing platform juggling visual perception, speech, and real-time motion control, with the compute split across robot, edge, and cloud tiers^[6][15]. The perception front end — cameras and depth sensors feeding detection, tracking, and pose — is where power, latency, and cost concentrate, and it is the part that maps directly onto near-sensor silicon.

Why now

China’s humanoid output is projected to grow 94% in 2026, led by Unitree and AgiBot^[7], and Beijing treats embodied AI as an industrial priority^[8]. Korea is mobilizing in response: Samsung raised its Rainbow Robotics stake to 35% and created a robotics office^[9][10], the major groups plan factory deployments from late 2026, and a 1-trillion-KRW government program funds on-device AI chips for robots, vehicles, and drones^[5]. Both ecosystems need exactly what a near-sensor chip company sells: perception within a power budget, without a cloud round trip.

What UXF brings

We are not building a robot. We co-design the perception subsystem — sensor, silicon, and model tuned together for the robot builder’s constraint, whether that is battery life, joint-loop latency, or unit cost. Measured energy per inference on the builder’s own model is the deliverable.

In July 2026 we signed an MOU with Everybot, the KOSDAQ-listed service-robot maker, covering joint development of humanoid AI semiconductors, robot vision and sensor-fusion edge AI, and commercialization of AI-semiconductor-based robot systems^[16]. Discussions with robot builders in China are opening.

MOU signed: Everybot (Korea) / China discussions opening

How we engage

Bring us the robot’s power envelope, its control-loop deadline, and the perception model you already run. We return a co-designed sensing front end with measured numbers against those three constraints. Request a briefing to start.