Cerebellum

From the brain to the algorithm. The cerebellum is the silent specialist: 80% of the brain's neurons in a structure the size of a fist, dedicated to making movement smooth, balance reliable and timing predictable. Its AI counterpart — humanoid robotics — is in 2026 in the same place computer vision was in 2014: capable, expensive, and approaching its breakthrough year.

What the biology does

The cerebellumW sits at the back of the head, under the occipital lobes. Anatomically small but computationally enormous, it holds roughly 80% of the brain's neurons in a tightly folded sheet.

• Procedural learning — riding a bicycle, touch-typing, throwing a ball. Each is a learned forward model.
• Fine motor coordination — the smoothness of a violinist's bow.
• Predictive forward models — anticipating where your hand will be in 50 ms based on the motor command you just sent.
• Balance and timing — sub-100 ms feedback loops that classical control engineers envy.

Damage produces ataxiaW: intent is preserved, execution is not. You know what to do; your body does it badly.

What we have built

Humanoid roboticsW finally reached commercial relevance between 2023 and 2026. Twelve milestones, four production-grade deployments, one inflection year.

• 2021 — RL-based legged locomotion. ANYmal (ETH Zürich) and Boston Dynamics Spot demonstrate production-quality reinforcement-learnedW walking.
• July 2023 — Google RT-2. First vision-language-action model — a Transformer that emits robot motor tokens trained jointly with web image and text data.
• 2023 — Humanoid prototypes. Tesla Optimus, Figure 01, 1X Neo and Unitree H1 reveal commercial humanoid platforms aimed at general-purpose work.
• October 2024 — Physical Intelligence π0. First open-source generalist robot foundation model.
• March 2025 — NVIDIA Isaac GR00T N1. First open humanoid foundation model + the Newton physics engine for sim-to-real.
• April 2025 — Physical Intelligence π0.5. Open-world generalisation: π0.5 cleans new kitchens it has never seen.
• 2025 — Figure 02 at BMW Spartanburg. Logs 90,000+ parts and 1,250 hours across more than 30,000 X3 bodies — the first humanoid production case study at scale.
• October 2025 — 1X Neo pre-orders open at $20K. First consumer humanoid for the home — 1X Neo ships with 22-DoF dexterous hands and a 22 dB noise floor.
• November 2025 — Agility Digit passes 100,000 totes moved at GXO. First commercial humanoid milestone in a live warehouse.
• January 2026 — Tesla Optimus Gen 3 production deployment — 1,000+ units on Tesla factory floors.
• February 2026 — Apptronik $935M Series A; Apollo at Mercedes, Figure-03 at BMW. The European pilots cross from press release to payroll.
• 2026 — BMW Group AEON in Leipzig. First humanoid in a German plant — full pilot summer 2026.

The architectural recipe arrived in mid-2023 and has not been displaced since:

"Our extensive evaluation (6k evaluation trials) shows that our approach leads to performant robotic policies and enables RT-2 to obtain a range of emergent capabilities from Internet-scale training. This includes significantly improved generalization to novel objects, the ability to interpret commands not present in the robot training data […], and the ability to perform rudimentary reasoning in response to user commands." — Brohan et al., 2023 (arXiv:2307.15818)

Every commercial humanoid foundation model since — π0, π0.5, GR00T N1 — descends from RT-2's central trick: action tokens treated like language tokens, trained jointly with web-scale data. The Stanford AI Index 2026 tracks the resulting capability curve; Meta's V-JEPA 2 world model adds a physical-prediction prior on top.

What is still missing

Locomotion is in good shape. Manipulation, generalisation and learning efficiency are not.

1. Dexterous in-hand manipulation. Unscrewing a jar, threading a needle, separating two sheets of paper — current humanoids are far below toddler level on unfamiliar objects.
2. Sample efficiency. Robots learn in millions of simulated episodes. A child learns to button a shirt in tens of attempts.
3. Robustness on the real world. Uneven terrain, contact-rich environments, deformable materials and unstructured clutter still break policies that work in the lab.
4. Generalisation. Most humanoids excel at the narrow tasks they trained on and fail on the next one over.

How we read the verdict

We rate the AI counterpart Developing. The progress curve is steep, the commercial pull is intense, and the architectural recipe is converging on transformer-based VLA models. We do not have a cerebellum yet, but for the first time it is plausible that we are within a decade of one.