NEURA Robotics Unveils Neuron AI Foundation Model for Physical AI

NEURA Robotics introduced Neuron AI, its multimodal foundation model for Physical AI, on October 14, 2026, in an announcement from Metzingen, Germany. According to the company, the model’s Perception-Language-Action (PLA) architecture ties a robot’s sight, hearing and touch to what it has already done, converting multisensory information into physical action.

Founder and CEO David Reger said intelligent robots need a brain that unifies their senses and remembers past events. “With Neuron AI, we are building the brain of the physical world,” he said, describing the model as Physical AI developed in Germany and built for the world.

The product page breaks the model into three parts: a perception layer through which the robot sees, touches and hears, knows its own body and takes the user’s command; a language model that brings every input into a shared language space and reasons over the request; and an action model that translates decisions into precise, safe robot motion. In that pipeline, every modality is encoded into a common, time-aligned representation before being conditioned into one stream of action.

Multisensory Perception and Memory

NEURA’s case for the model starts from the limits of cameras: they capture geometry, color and position, with no information about contact, force or sound. People, the announcement notes, do not work by sight alone: they feel a screw catch and hear a connector snap into place, and Neuron AI applies the same principle to robots.

The company said the extra senses matter most for tasks cameras cannot solve alone. When a robot opens a lock, only sound and touch confirm it is really open, and touch lets a robot grasp objects hidden from view. They also sharpen tasks vision already handles: in welding, the sound of the process can indicate the quality of a seam.

The model also carries a memory function. NEURA argues that acting in the physical world requires a sense of cause and effect and object permanence — the ability to remember objects when they leave view, which humans acquire in their first two years of life. Unlike purely digital AI, the company said, a robot acts in a continuous physical world and must track objects and the consequences of its own actions over time, often while not seeing them. NEURA said it reshaped the architecture specifically to introduce memory and trains it on dedicated datasets built to use it.

On the product page, the company contrasts vision-led vision-language-action (VLA) models, which it says are optimized for motion and reason mostly on the current frame from what the camera sees, with Neuron AI, which it says covers the same ground as a VLA while adding touch, sound and memory to cover what cameras miss.

Three Data Pillars and Five Embodiments

Neuron AI is trained on three complementary sources of physical experience. Robots of different embodiments work through tasks in the NEURA Gyms, the company’s training infrastructure for cognitive robots and Physical AI, and in real-world settings, producing synchronized streams of vision, touch, force, audio and movement data. NEURA’s Data Amplifier and simulation stack broaden that data with simulated variations of tasks and environments, and human demonstration data comes from people wearing sensor-equipped data suits, whose recorded movements and task demonstrations carry practical expertise into training.

The three pillars supply real physical interaction, simulated variation and human know-how, training one shared foundation across five NEURA robot embodiments with different grippers and hands, the company said. The product page identifies the five as 4NE-1 Gen 3, 4NE-1 Gen 4, a wheeled 4NE-1 and two further bimanual platforms, and lists MAiRA, MiPA, MAV+ and 4NE1 as the robot family across which one model is trained and deployed.

On-Robot Deployment and the Neuraverse Platform

Neuron AI runs locally on the robot itself, the product page states: decisions are taken in real time on the robot, with no server round trip, it continues operating even without a cloud connection, and production data stays under the customer’s control, which the page calls a key point for manufacturers in Europe.

Within the NEURA Gyms, data collection, skill training and benchmarking are connected ahead of deployment, and the data and skills produced there feed into later model releases. The Neuraverse platform then deploys and updates these capabilities across robot fleets and gives customers tools to fine-tune the model for their own tasks, so that a customer’s process expertise becomes robot skills. Skills can remain private to the customer; where a customer shares skills or deployment data, NEURA said, those contributions can support the wider ecosystem and future model development.

The company makes Time-to-Skill, the speed at which a robot learns a new task and performs it reliably, a central measure of progress. For industrial customers, NEURA said, shortening that time is essential to expanding Physical AI into more processes and adjusting it as requirements change.

German Development and the GTC Berlin Demonstration

Development and training took place in-house at NEURA’s Metzingen site, where the company brings together the robots, hands, sensing technology, data infrastructure and AI development needed to tie model training to physical performance. NEURA said Germany’s industrial base of complex manufacturing processes, specialized expertise and demanding operating environments provides an important setting for this work, and data partnerships with industrial companies including Schaeffler and Bosch add to the training carried out in the NEURA Gyms.

At NVIDIA GTC Berlin 2026, NEURA will demonstrate its humanoid robot 4NE1 (“For Anyone”) running the current version of Neuron AI through an eleven-minute laundry-handling sequence. In that sequence, 4NE1 sorts laundry, loads a washing machine, closes and reopens it, unloads it and moves the laundry to a dryer, all without human intervention. The company said the difficulty lies beyond any one movement: the robot must hold relevant context across the whole sequence, finish each step and move on to the next. The demonstration is scheduled at Station-Berlin, Hall 3, Booth #3031, from October 20 to 22, 2026.