Praxis One Brings Robot Decision-Making and Motion Control Onto One Chip

Huaruizhipu introduced Praxis One during China Robot Summit week in Hangzhou on September 17, 2026. Built around MediaTek’s Genio Pro 5100 system-on-chip, it is the higher-compute version of the company’s on-robot physical-AI platform, aimed at heavy-duty humanoids and other robots with demanding perception and movement workloads. 710

Why put the robot’s “big brain” and “small brain” on one chip?

Huaruizhipu’s design brings cognitive decision-making and motion-control computing onto a single chip. In a split setup, data must pass between a cognitive accelerator and a separate controller. The company says its approach can reduce that communication delay, data-transfer overhead and the work of adapting two hardware systems to each other. Keeping more processing on the robot also avoids relying on a cloud round trip for every local decision. 71017

That is an architectural rationale, not a measured latency result. The cited reports do not establish a Praxis One control-loop time or show that it outperforms a particular split system in a robot deployment. 717

What workloads is Praxis One meant to handle?

Praxis One is positioned for multimodal perception, vision-language-action (VLA) models, local language-model inference, lightweight world-model workloads and complex task planning. Reports also describe the Genio Pro 5100 CPU as suited to demanding kinematics and planning in changing environments. Together, those capabilities are intended to support a shorter path from sensing and inference to motion decisions on high-degree-of-freedom robots. 7104

The quoted Genio Pro 5100 specifications are substantial: a TSMC 3 nm process; eight Arm v9.2 CPU cores comprising one Cortex-X925, three Cortex-X4 and four Cortex-A720; 12 MB of L3 cache; 10 MB of system cache; and up to 260K DMIPS. Reports put its eighth-generation NPU 890 at 106 TOPS and describe a Transformer-focused engine. They also quote local Qwen3-7B generation at 23 tokens per second. 345

These figures describe the chip or a reported model run—not simultaneous performance across all of Praxis One’s proposed workloads. In particular, text-generation speed is not a motion-control benchmark. MediaTek separately describes over 50 TOPS of system-level generative-AI acceleration, a differently framed figure that should not be treated as interchangeable with the reported 106 TOPS NPU rating. 1413

Praxis One versus Praxis Nano

The distinction is workload and robot size. Praxis One uses the Genio Pro 5100 for heavier perception, model inference and planning on large or high-compute robots. Praxis Nano uses the Genio 720 and emphasizes real-time motion control, synchronized sensor collection and local feedback loops for smaller robots and lightweight data-collection devices. MediaTek lists the Genio 720 NPU at 9 TOPS, though chip ratings alone do not compare the two finished platforms’ control performance. 71012

Where it fits in Huaruizhipu’s wider plan

Huaruizhipu presented the two compute platforms alongside its Ruimou Xingyuan data-collection ecosystem, a physical latent-space world model and robots for specific scenarios. The company’s proposed turnkey direction is to combine computing, software, data tools and robot applications rather than offer a chip alone; the cited coverage does not establish how much of that integrated offering has been delivered to customers. 71031

Chinese coverage calls the overall platform “智能超脑”—literally “intelligent super-brain”—and identifies Praxis One and Praxis Nano as its two products. The available sources do not establish a definitive official English name for the overall platform. 710