Plug-in decision module
Plug in a card. Give your robot judgment.
Preloaded decision model, safety veto rules, and site packs — connect a camera and ROS 2 and run, with no cloud connection required.
Plugging in the accelerator is where the real work starts
Accelerator cards ship empty
You still have to build the model, quantize it, and write the safety logic yourself — most small and mid-size robot makers don't have the team for that.
Cloud brains don't fit every site
Cloud LLMs add latency, drop connections, and send your data off-site — often not an option for hospitals, factories, defense, and security-sensitive government sites.
Small robot makers don't have an AI team
You want smarter behavior, but building a decision-and-safety stack from scratch — and maintaining it — isn't realistic for most teams.
Three ways to start, one decision backbone
Whether you run it as shipped, bring your own site pack, or bring your own model, the same perception-to-action backbone runs underneath.
-
1
Ready to run
Plug in the module, connect a camera and ROS 2, load one site pack, and you're running the same day.
-
2
Bring your site pack
Describe your site — POIs, zones, elevators, skills — with our site pack authoring tool. No code changes needed.
-
3
Bring your own model
Swap in your own quantized model. The safety veto layer and ROS 2 interface stay the same.
-
4
Camera feed comes in
An existing camera or RTSP stream feeds directly into the module.
-
5
Edge VLM and decision model
The edge VLM flags site hazards — spills, caution signs, smoke, out-of-service notices — first, then the decision model proposes the next action.
-
6
Safety veto on a separate MCU
Proposed actions go to a separate, functional-safety-certified MCU for a veto check. Unsafe actions are blocked there, not left to the main decision model to police itself.
-
7
ROS 2 action and decision log
Approved actions go to ROS 2 for execution, and every decision is written to a log you can replay later.
What you get
Preloaded module
A module from one selected supplier — for example Hailo-10H or a Jetson-class board. The exact model depends on your site's power and compute needs.
Quantized decision model
The decision layer's enforcement contracts and site pack logic, quantized and ready to run on-device.
Independent safety veto MCU
Safety veto logic runs on a separate, functional-safety-certified MCU — it doesn't rely on the main decision model policing itself.
ROS 2 package
Connects to your robot's existing ROS 2 / Nav2 interface to pass navigation goals and action commands.
Site pack authoring tool
Describe POIs, zones, elevators, and skills for a site — no code rewrite when you move to a new one.
Model and site pack updates
Decision models and site packs keep improving without swapping hardware.
Who it's for
Small and mid-size robot makers
You want smarter features but don't have an in-house AI team to build decision and safety logic from scratch.
Plug in the module and get decision-making with safety veto and decision logs included.
Sites that can't connect to the cloud
Hospitals, factories, defense, and security-sensitive government sites often can't send data off-site or wait on a cloud connection.
Decisions run entirely on-device — no network connection required.
Accelerator vendors who want to sell solutions
Your chip sells compute, but customers still have to add the model, quantization, and safety logic themselves.
Ship your silicon as a complete decision module — a new product position, not just a chip.
Why us
The accelerator modules and robot-brain platforms on the market are solving a different problem.
Others sell compute. We sell judgment.
Modules like Hailo and Axelera, or NVIDIA Jetson, sell raw compute. Plug one in and it's still empty — you still have to add the model, quantization, and safety logic.
Robot-brain platforms target large players
Platforms like Jetson Thor and Qualcomm Dragonwing IQ10 are built mainly for large companies and humanoid robots with their own AI teams.
Every decision has a safety check and a log
Our safety veto runs on a separately certified MCU, not on the decision model policing itself — and every decision is written to a replayable log.
Licensed, not sold once and forgotten
Modeled on Brain Corp's approach: software licensing plus a certified controller. Hardware price includes the license; updates are a separate fee, so we stay accountable for what ships after day one.
Pricing structure
Evaluation kits are reservation-only and ship once our edge benchmarks are complete. Production volumes and custom site packs are quoted.
Evaluation kit (reserve)
The starting point — reserve a module now, receive it once edge benchmarks are done, and test it against your own site.
- Preloaded module with software license
- Site pack authoring tool
- Direct technical support
Reservation only. Not shipping yet; timing depends on edge benchmark results.
Production license
Per-unit licensing with model and site pack updates.
- Hardware price includes software license
- Annual model and site pack update fee
- Onboarding and custom site pack service
Volume and custom site packs quoted separately.
Advanced subscription
Modeled on Brain Corp — a per-unit, per-term subscription for advanced decision and safety features.
- Advanced decision model subscription
- Expanded site packs and skills
- Decision log review and replay
For customers who need ongoing updates and expanding site coverage.
Current status and validation roadmap
Verified
- The decision layer passes 435 automated tests, including site packs and decision log replay
- 4 site packs are ready to reuse
- The edge VLM (Cosmos3-Edge) identifies four hazard types on simulated imagery: spills, caution signs, smoke, and out-of-service notices
In progress
- Decisions currently run on a cloud-scale model; moving to a model small enough for an edge chip may reduce decision quality, and that's still being verified
- A model like V-JEPA 2 takes about 16 seconds per action on an RTX 4090; it will be slower on edge modules with far less compute, and no one has published edge-chip numbers for this yet
- ROS 2 / Nav2 integration is verified in simulation only, not yet on a physical robot
Next
- Benchmark one decision's latency and quality gap against the cloud on both Jetson Orin Nano Super and Hailo-10H
- Plan production, inventory, and functional-safety certification
- The accelerator supplier market is consolidating fast — weigh the lock-in risk before committing to a first supply partner
FAQ
Can you really make decisions without the cloud?
Yes — the decision model and safety veto both run on the module. Edge decision quality is still being verified; see the status section above.
Is the safety veto software or hardware?
It runs on a separate, functional-safety-certified MCU, independent of the main decision model — it doesn't rely on the model policing itself.
Can we swap in our own model?
Yes — bring-your-own-model is one of three supported setups. The safety veto layer and ROS 2 interface stay the same.
Which cameras and ROS 2 versions do you support?
That's confirmed during the pre-deployment site assessment — tell us your camera specs and ROS 2 version in the form below.
How do updates work?
Models and site packs keep updating, covered by the annual fee — no hardware swap needed.
Why not just buy a Jetson and build it ourselves?
You can — but you'd still have to quantize a model, write the safety veto logic, wire up ROS 2, and log every decision. We've already built that layer and shipped it inside the module.
Let's look at your site
Tell us about your robot, your site, and what you're trying to solve. We'll set up a conversation or send evaluation kit details.