ODM / OEM
AI companion robot development and manufacturing
Desktop, elderly-care and children’s companion robots. The part that decides whether one ships is not the character design — it is response latency, wake-word reliability in a noisy room, and a battery and thermal budget that survive a closed shell.
Where people usually are when they call
You have only the requirement
No demo, no software, no drawing — just a clear idea of what it should do and who it is for. We take it from there: industrial design, mechanical, hardware, embedded, app and server. You define the product; we build all of it.
You have a demo
A dev board on a desk that talks back. The distance from there to a manufacturable unit is the silicon, the structure and the certification — which is the work we do.
You have the software
Your model integration, your prompts, your app and backend. They stay yours — we adapt the device to them rather than replacing what already works.
You have the channel
You know who buys it and at what price. You need a unit that exists, certified for your market, at a cost that leaves you a margin.
What actually decides a companion robot
Response latency
Under roughly 700 ms after the user stops speaking it feels like conversation; past 1.5 s it feels broken. With the model in the cloud, most of that budget is spent on the device before the network is touched — wake word, endpointing, noise suppression, speech recognition. The NPU, the audio front end and the streaming design have to be decided together.
Wake word in a real room
A wake word that works in a quiet office fails next to a television. Microphone array geometry, enclosure port design, echo cancellation while the speaker is playing — these are mechanical and DSP problems as much as software ones.
Motion makes noise and heat
Servos for a moving head or body are heard by the microphones and felt as vibration. Compute makes heat that a sealed toy-sized shell cannot shed. Both are solved in the structure, not in firmware.
Offline behaviour
The network will drop. What the device does then is a product decision you have to make early, because it determines whether you need local speech recognition or a small on-device model — and that changes the silicon tier and the cost.
Children and privacy
A companion aimed at children brings EN 71, CPSIA, capped audio levels and a parental-control story. A camera brings a privacy story — a physical shutter answers more questions than a settings toggle.
Battery that survives certification
A lithium pack needs UN38.3 before anyone will air-freight it, and the capacity you want may move the product into a different shipping class. Decide this before tooling, not after.
Two architectures, and when each is right
Wireless SoC + cloud model
Wake word and audio front end on a Bluetooth/WiFi SoC; recognition and the model in the cloud. Lowest BOM, smallest board, longest battery.
Right when · Right when the product is always connected and the price point is tight.
Linux SoC with NPU
0.5–16 TOPS on the device: local speech recognition, vision, and from around 6 TOPS a quantised small model for offline fallback. More silicon cost, more capability, more heat to design around.
Right when · Right when it has a camera, must work offline, or the latency target is aggressive.
Questions we are asked
Can you work from our demo?
Yes — that is the most common way a project starts here. Send a video of it running and your target retail price; we come back with a silicon recommendation, a cost envelope and the certification path.
Do we have to use your app and cloud?
No. If you already have an app, a backend and a model integration that work, they stay yours and we adapt the device to them. We supply those layers only when you want them.
Which language models can it use?
Whichever you choose — bring your own API key or your own endpoint. The device handles wake word, audio front end and speech; which model answers is a configuration, and the cloud region is yours to decide.
What is the minimum order quantity?
Set per project, driven by tooling amortisation and the minimum order on the parts chosen — not by a number we publish. We quote it after the feasibility step, when the silicon and structure are settled.
Will you sell the product yourselves?
No. We build no finished-goods retail channel and carry no brand of our own. The design and the market stay yours; IP terms are fixed in the contract and an NDA is signed before any specification is exchanged.
How long from brief to production?
It depends on how much is new — a variant on a proven platform and a ground-up design are not the same project. We commit to a schedule after feasibility, and we would rather quote a date we can hold than one that wins the enquiry.
Request a Quote
Tell us what you are building.
A video of the demo, your target price and the markets you sell into is enough to start. An engineer answers within one business day.
Related