Any physical AI.
On any robot. Deployed safely.
Subcortic lets factories and robotics teams run modern AI models on whatever hardware they already own. Register a machine once, block dangerous movements before they reach the motors, roll out updates safely, and automatically retrain from real factory floor runs.
$27.6B
invested into physical AI startups in 2025, plus $16.3B in Q1 2026 alone
4.66M
industrial robots already sitting in factories worldwide waiting for AI
542K+
new factory robots installed every year across manufacturing and logistics
100%
of unsafe movements into hazard zones blocked before reaching the motors
$250
per robot per month for continuous safety, updates, and fleet retraining
Sources: PitchBook 2026 Physical AI Report and IFR World Robotics.
The bottleneck
$27.6B is chasing a better brain. Nobody is selling the nervous system.
Billions are going into AI models, not the factory floor
AI labs are building impressive general-purpose models, but none of them ship the software needed to make those models run reliably on a customer's actual factory machines.
Every robot on the floor is different
A single facility runs a mix of industrial arms, older retrofits, and custom cells. Rewriting custom integration and control code for every machine takes months and breaks on every update.
One bad AI prediction wrecks a $50,000 machine
AI models are probabilistic. Without a hard, physics-based safety layer between the model and the motors, no plant manager can risk running learned models near people or expensive tooling.
How Subcortic works
Register once. Deploy safely. Improve every week.
Subcortic turns raw AI models into dependable factory workers. Instead of building custom software for every machine, you plug any robot into one continuous deployment and learning loop.
1. Register
Define any robot once: dimensions, speed limits, and off-limits hazard zones
2. Adapt
Fine-tune any AI model to your specific hardware from a handful of demos
3. Gate
Automatically block any model update that fails accuracy or safety stress tests
4. Deploy
Roll out verified updates to one test robot first before updating the whole fleet
5. Improve
Capture real factory runs and automatically retrain so the fleet gets smarter every week
Universal Robot Profile
One simple configuration file adapts the platform to any industrial arm, retrofit, or custom machine without custom code.
Pre-Release Stress Gate
Every model update is tested against sensor noise and edge cases before it is allowed near a real robot.
Real-Time Safety Shield
A physics-based safety layer checks every command in milliseconds and stops the robot before any part of it can hit an obstacle.
Closed-Loop Fleet Retraining
Robots automatically upload their successful work runs so your models continuously improve on the job.
Built for the real world
An AI model never touches a physical motor unsupervised.
Safety cannot be another AI model that guesses. Subcortic enforces hard, predictable physics rules between the AI and the machine on every single millisecond of movement.
Full-body collision protection
Most software only checks where the robot's gripper is going. Subcortic tracks the entire physical body of the robot, including every joint and elbow, so no part of the machine can ever swing into a worker station or cabinet.
Automatic braking distance
A heavy robot arm moving fast cannot stop in zero millimeters. Subcortic calculates the exact physical stopping distance on every step and slows or stops the machine before momentum can carry it into a hazard zone.
Protects gears and motors from AI spikes
When an AI model gets confused, it can command sudden violent jerks that strip gearboxes. Subcortic smooths out sudden spikes and prevents the robot from locking up in awkward positions.
Mandatory pre-deployment stress test
No model update goes straight to the floor. Every new version must pass automated tests under simulated sensor glitches with zero safety violations before it is approved.
Tamper-proof updates
Every model release is cryptographically locked to the exact robot model it was built for. A corrupted file or an update meant for a different machine is rejected automatically.
Canary rollout and instant rollback
New updates always run on a single test robot first. Only after that robot proves itself on the floor does the update roll out to the rest of the fleet, with one-click rollback at any time.
Built to align with ISO 10218-1:2025 industrial robot safety standards.
Business model
Start with one robot. Scale to thousands of recurring subscriptions.
Setting up one robot is a project. Keeping hundreds of factory robots safe, updated, and improving every week is permanent software infrastructure.
Workcell onboarding
$15K to $40K one-time
We onboard a customer's first robot cell, configure its physical safety boundaries, and deploy their first working AI policy in days instead of months.
Per-robot subscription
$250 / robot / month
Recurring software license for every active robot on the floor: real-time collision protection, staged over-the-air updates, and compliance audit logs.
Automated fleet retraining
Usage-based compute
As robots work on the floor, they capture new task data and automatically retrain in the cloud, turning every customer fleet into a compounding data asset.
A $1.1B+ recurring software market on existing robots alone
There are already 4.66 million industrial robots operating worldwide, with over 540,000 new units installed every year. At $3,000 per robot per year, capturing just 8% of the existing installed base represents over $1.1B in annual recurring revenue.
Why Subcortic
Why factories can't just plug a raw AI model straight into a robot.
AI model labs build great general-purpose intelligence, and traditional robotics tools build great dashboards. Subcortic bridges the gap between them so AI models can actually work on a factory floor.
Ready to deploy physical AI on real machines?
Explore our live interactive simulator and fleet control plane, or read the investor pitch deck.