On August 19, during the World Robot Conference, Huixi Intelligence released its Huixi Embodied product series. The launch focused on chips, modules, a software development environment, and engineering deployment — the robots' physical movements were only one part of the story.
From Chip to Workstation
The R1 PRO is an embodied-scenario SoC fusing the "brain" and "cerebellum": it places CPU, NPU, GPU, and real-time interfaces on a single chip, allowing perception, reasoning, decision-making, and motion control to close the loop on-chip. Official specifications include:
- 375 TOPS INT8 AI compute
- 16-core A78AE CPU
- 1 kHz real-time control
- 204.8 GB/s LPDDR5 bandwidth
- Up to 22 camera inputs
- Robot-common interfaces: Ethernet, CAN FD, PCIe, USB
Above the chips sit the RCM 400 and RCM 500 compute modules. RCM 400 pairs with R1 PRO, offering 16GB, 32GB, or 64GB LPDDR5; RCM 500 pairs with R1 MAX, offering 32GB, 64GB, or 128GB. Both use a 699-pin interface and the RISE / Robotics SDK. The modules absorb repeated design work — memory, UFS, power management, clocking, and high-speed signals — so robot companies don't have to start from a multi-layer PCB each time.
What WRC Says About the Industry
First-day WRC coverage noted over 200 embodied intelligence exhibitors this year, with 300+ physical robots running on-site; exhibits shifted from backflips to guided tours, logistics handling, production-line assembly, and inspection. The conference also introduced a "Procurement Day" for the first time. These are media observations, not a specific company's order book, but they explain why Huixi chose this moment to launch a full-stack platform: customers have started asking about delivery cycles, interface stability, and ongoing operating costs — the questions have changed.
Huixi also released an AI-augmented development kit, attempting to let agents participate in model deployment, performance profiling, operator development, and tuning. The company's stated goal: cut deployment cycles for the same algorithm from 2 weeks to 2 days. This claim should be read carefully. Two weeks to two days is a launch target, not publicly available third-party measurement; what really needs verification is under which models, sensors, and power constraints the reduction holds, whether operators must be rewritten, and whether field failure rates actually drop.
One SoC Isn't Enough
TOPS is a peak metric and doesn't directly represent a robot's success rate at grasping, sorting, or assembly. Hard constraints in embodied systems include sensor time synchronization, control-loop jitter, memory bandwidth, thermal management, interface latency, and software migration costs. R1 PRO's value in putting these on one chip lies in reducing multi-chip communication and board-level integration complexity; whether it enables skill reuse across different robot bodies still depends on whether the SDK, model formats, and training data are genuinely shared.
This distinguishes the news from earlier narratives like "the same brain across bipedal, wheeled, and heavy-load frames." Galbot emphasizes generalization at the model and body layers; Huixi is building a platform out of the compute substrate and engineering interfaces. The two are adjacent in the value chain but do not replace each other.
For embodied intelligence to cross the "demo works" threshold, the next gate comes down to plain questions: how much code changes when one algorithm moves to another robot, how many manual restarts a week of continuous running requires, and who maintains low-level drivers during production-line upkeep. Chip launches tend to hide the answers in a spec sheet — WRC's Procurement Day pushes them back into the open.
Sources and verifiable links
1. 36Kr: Huixi Intelligence releases Huixi Embodied series (2026-08-19) https://36kr.com/p/3946324792556672 2. 36Kr: WRC 2026 day-one observations (2026-08-20) https://36kr.com/p/3946697386808967