Learn is how Inztinct improves. Robots learn from their own sites, share what they learned without sharing raw footage, and receive improvements only as tested, signed releases. Learn can change Sense and the planner. It has no path to Core.
Two steps are gates. If either one fails, the release stops.
Each robot logs near-misses, stops and corrections. Raw data stays on the machine or on the site's own server.
Candidate updates are trained away from the robot. A running robot never learns live.
Only model updates leave the site (federated learning). Open frameworks such as NVIDIA FLARE are an option we are evaluating.
Odd or poisoned updates are rejected before they are merged.
Every Core limit and every recorded scenario is replayed in simulation. One failure stops the release.
In every scene tagged hazardous, the new model must allow no more than the approved baseline. It may relax caution elsewhere, which is how nuisance stops come down.
A person approves. The model is frozen and signed. The robot installs only signed releases.
A few robots first, with rollback ready and the logs watched.
A ward, a factory floor or a mine pit does not need to send raw video away for the fleet to learn from it.
Training and checks run offline. Releases move when a link is available, or by hand.
A lesson from one corridor can improve every robot in the fleet, once it passes the gates.
A bad site or an attacker sends updates meant to teach the fleet something unsafe. We check contributions and the gates catch behaviour changes in hazard scenes.
Model updates can reveal something about the data behind them. We treat updates as sensitive and are evaluating protections.
A model that suits one site may suit another worse. Staged rollout and rollback exist for this.
These are the risks we plan for. None of the protections are built yet.
The claim we want to be able to make: robots across sites got better, and no release failed the safety suite. We will publish the numbers once there are numbers.
Core holds the line. Sense reads the room. Learn gets better.