Back

Event-based

Our world is messy and nature figured out the best way to understand it is via changes. That is how The Minimalist sees and understands the world.

Continual learning

The next breakthrough in physical AI is not stuffing data down a hyper-scaled transformer. It is by learning how the most efficient computer in existence learns — a.k.a. biological brains.

The Minimalist

We built a neurocognitive model from the ground up, using neuro processes that we turned into learning algorithms. This is neuroscience turned intelligent.

More from less.

13% zero-shot generalization from 5.2M samples — no policy on the leaderboard learns to generalize faster.

Scatter plot of generalization rate (percentage points per million training samples) versus total training samples for twelve policies
Composite-Unseen · zero-shot success rate

RoboCasa365 leaderboard, Composite-Unseen split, retrieved 07/21/2026 — robocasa.ai/leaderboard.html. Training samples = training steps × batch size (sample presentations during training). Per the leaderboard’s own note, training configurations differ across architectures and step counts are not directly comparable.

Composite-Unseen success rate, training samples, and generalization rate by policy
PolicyComposite-Unseen %Training stepsBatch sizeTraining samples (M)Rate (pp per 1M samples)