The controller for an embodied robot — the hardware is done.
Motion control is what's missing.
For motion-control FDE talent across Shenzhen and Hong Kong — this is the layer you'd own.
The Shenzhen–Hong Kong AI Application Innovation Centre provides controller hardware for embodied robots — currently the X-3321, built to serve as a robot's "brain and cerebellum" in one box: twelve automotive-grade camera inputs, 275 TOPS of on-device inference, four EtherCAT channels plus CAN-FD / RS485 / DIO running straight down to the servos and actuators, on 9–36V so it mounts directly on the machine. That half is complete. What is missing is the motion-control software layer. We are looking for outstanding FDE (forward-deployed engineer) talent across both cities to build it with us, and to get it running on a real machine in a real setting.
What an embodied robot lacks is never compute — it's the layer that makes it move
For an embodied robot to move, two layers are needed: perception and decision-making (the brain), and real-time motion control (the cerebellum). Both have to land on the same controller — one that takes multiple camera feeds, has the compute for on-device inference, drives actuators in real time, and still mounts on a machine that moves, vibrates and runs on a battery.
Controllers that meet all of those at once are not common. But the hard part isn't the hardware — in a technical specification for an embodied-robot cerebellum controller, of thirty-odd requirements only a handful land on hardware. Everything else has to be implemented in software.
We have finished that half — the X-3321 was built to exactly this brief. What is missing is the motion-control software layer, which decides whether the controller ever becomes a machine that actually moves. Shenzhen and Hong Kong are not short of people who can build that layer; what is short is hardware they can genuinely try it on.
The Centre provides the controller, bench space in Shenzhen and engineering support, and looks for motion-control FDE talent across both cities to build that layer and get it running on a real machine in a real setting. Hardware and engineering on our side; motion control and deployment on yours. Once it runs, how we keep working together is something we talk through in person.
Jia Heng Technology Development Co., Ltd. is the operating entity, registered in Hong Kong and operating across both cities: the Shenzhen side holds the hardware, the bench and the supply chain; the Hong Kong side is connected to the universities and to international capital. Take away either side, and this stretch of work doesn't get done.
Motion-control FDE talent in Shenzhen and Hong Kong — we're looking for you
Only what we can actually deliver. Places are limited and dated, so that what we say yes to gets done.
Right now, it's the X-3321
It isn't the only one — it's the one you can put your hands on right now. It was built to be an embodied robot's brain and cerebellum in one box: twelve automotive-grade camera inputs, 275 TOPS of on-device inference, four EtherCAT channels plus CAN-FD / RS485 / DIO running straight down to the servos and actuators, 9–36V and 1.72kg so it mounts on the machine itself.
Build the layer the X-3321 is missing: motion control
You might already work as an FDE, or you might be headed that way. What we're judging isn't your résumé — it's whether you can make it actually move: an EtherCAT master (IGH / SOEM), a 1 kHz control cycle under PREEMPT_RT, multi-axis DC synchronisation, position / velocity / torque loops, impedance and admittance control, joint- and Cartesian-space interpolation, inverse kinematics, safety limits and watchdogs. Whichever of those you know well is where we start.
- One X-3321 controller, on our bench in Shenzhen or taken back to your own space.
- Four switchable power modes: 15W / 30W / 50W / MAXN — on a battery-powered machine this converts directly into runtime. Module-level power read directly, whole-unit power via an inline meter.
- Bench space and compute in Shenzhen, plus an engineer working alongside you for two weeks.
- Capability in the motion-control layer — EtherCAT master, real-time kernel tuning, multi-axis synchronisation, hybrid force-position control, trajectory planning and inverse kinematics. Having done any one of these is enough to start the conversation.
- One specific motion you want to make happen, and the willingness to spend the time getting it running, measured, and documented so someone else can redo it.
- A one-page proposal covering three things (next section).
- A motion-control prototype running on real hardware — a closed loop from perception to actuation, with a power-and-latency table across the four modes.
- A reproducible script and environment note that someone else can re-run and match.
- The method, code and conclusions are yours — publish it, put it in your portfolio, either works.
What we can offer this round is the hardware, the bench and engineering hours — not an employment contract. How far this layer gets, and whether we keep working together, is something we figure out once it moves.
Starts mid-October, runs four weeks. Two places, closing 30 September.
Send us a one-page proposalFrom algorithm to prototype
You get: engineering and supply chain in Shenzhen — enclosure, thermals, power, component selection and small runs. A hardware revision takes weeks, not quarters. We ask: an algorithm that already runs, and one specific thing you want it to become.
Course and degree projects plug straight in
For universities: capstone, final-year and course projects can run directly on the Centre's hardware and data — we supply the machines, real scenario data and company attribution, with timelines following the academic term and scope fixed on one page before the project starts. What never moves: supervisors keep academic direction and publication rights, and a student's graduation and publications are never subject to commercial terms.
One page, three things
No résumé template, no results needed in advance. Whether you're coming from academia or industry, just write directly — we welcome different approaches to the same motion.
-
Which motion you want it to make
Down to one motion. "Do motion control for embodied robots" does not count; "run an IGH master on the X-3321, hold six-axis DC synchronisation jitter to microseconds, and keep the control cycle steady at 1 kHz" does.
-
How you'll know it worked
On what machine, at what control rate, within what latency ceiling, and what counts as a failure. Rough is fine; reproducible is not optional.
-
What you need us to open up
The controller, the cameras, bench space, compute, or an engineer's time. Say what you need on the machine side too — whether you have one, or we need to find one together.
We reply within two weeks. If it fits, the next step is half a day together in front of the machine — not a signed agreement.
Privacy
Where personal data is involved, collection is minimised, processing stays on the device wherever possible, and participants may withdraw at any time. Where the work runs through a university project, it goes through university ethics review first; otherwise the Centre carries out its own privacy assessment. Anything crossing the border follows each jurisdiction's rules separately.
Publication
Methods and conclusions are publishable. Where participant identity or personal data is involved, one desensitisation review applies, with a 15-working-day window; no response means cleared. Review is never a lever over your conclusions.
IP
The method, code and conclusions you produce belong to you (where the work runs through a university project, academic outputs and thesis rights belong to the student and the university, as usual). The Centre claims no exclusivity over general methodology; only where the work combines directly with an existing product to form a patentable result do we agree separate terms on joint ownership and licensing.
What you build is yours. Where students take part, their graduation and publication are never subject to commercial terms. That is the premise, not a concession.
These we do not have
Naming the gaps is what lets everything in the section above stand up.
A robot to put it on
We build the controller, not the machine — wheeled or legged bases, joint modules, robotic arms: we have none of these ready to hand over. To validate on a real machine, either you have one on your side or we find one together. This is here because it directly determines how far you can get once you arrive.
Comparable numbers on a domestic NPU
Move the same method from Jetson to a domestic NPU and there is almost no public data on what accuracy, latency and power actually cost. We are working on that migration; the boards are not yet lendable. When they are, they move into the section above. It is also, in itself, an open question nobody has answered.
What we have been doing
Reverse chronological, and only things that actually happened.
-
First round of FDE recruitment opens. We provide the X-3321 controller plus bench and engineering support in Shenzhen, and are looking for motion-control FDE talent across both cities to build the layer the controller is missing. Two places, closing 30 September.
-
Our "open lab" column launches on Zhihu — edge-side test runs we have actually done, published in full: commands, environment, raw output and conclusions, so anyone can redo them.
-
Completed a full replication of SWS-Net (AAAI-26, WiFi sensing) with all eight ablation pairs matched, and discussed the discrepancies we measured with the paper's first author. Edge-side power and latency runs (Jetson Orin / domestic NPU) are in progress.
-
Attended ICML 2026 in Seoul under the shengangai academic identity, building contacts with researchers at several universities around on-device deployment.
-
The Centre completed its Hong Kong business registration under the name "Shenzhen–Hong Kong AI Application Innovation Centre", with a registered business nature of "AI frontier technology transfer and industry scenario co-creation platform".
-
Attended AAAI-26 in Singapore, following papers on edge inference and wireless sensing and talking with their authors.
One page is enough to start
Partnerships, proposals or questions — write anytime; we reply within two weeks.
lxy@shengangai.com