NEURA Robotics presented Neuron AI, a multimodal foundation model for its robots, on 14 October 2026. The German manufacturer, based in Metzingen, describes a system that combines 3D vision, touch, force data, audio, robot state and language, along with a memory function. The stated aim is to let a robot act on information that cameras alone cannot capture.
NEURA calls its architecture Perception-Language-Action (PLA). In practice, the model encodes several sensor streams, aligns them in a shared space over time and converts them into a single action stream. The manufacturer gives examples of tasks where a camera is not enough: to check that a lock opens, only sound and touch confirm its actual state; in welding, the quality of a weld bead can be judged by the noise of the process.
How Neuron AI’s memory tracks objects out of view
NEURA puts particular emphasis on the model’s memory. A robot acting in the physical world has to keep track of objects even when they leave its field of view, as well as the consequences of its own actions. The manufacturer says it adapted the architecture specifically to add this memory and trained it on dedicated datasets.
According to David Reger, founder and CEO of NEURA Robotics, “an intelligent robot needs a brain that brings its senses together and remembers what has happened.” He describes Neuron AI as a model developed in Germany and designed for the whole world.
Neuron AI training: in-house development with Schaeffler and Bosch data
Neuron AI was developed and trained in-house in Metzingen, according to NEURA. The manufacturer says it brings together the robots, hands, sensors, data infrastructure and AI development required. Training draws on three sources of physical experience and on data partnerships with industrial companies, including Schaeffler and Bosch, which complement data collected in the NEURA Gyms, the group’s data collection and training facilities. According to the press release, the model covers five NEURA robot models with different grippers and hands.
Deployment and updates go through Neuraverse, the manufacturer’s platform, which is also meant to let customers fine-tune the model for their own tasks. NEURA states that customers can keep their skills private, and that those they choose to share can feed into the ecosystem.
4NE1 humanoid laundry demo at GTC Berlin: an eleven-minute sequence
At NVIDIA’s GTC Berlin 2026, NEURA plans to demonstrate its 4NE1 humanoid robot carrying out an eleven-minute laundry sequence with the current version of Neuron AI. The sequence includes sorting laundry, loading a washing machine, closing it and opening it again, unloading it, then transferring the laundry to a dryer. NEURA says the sequence runs without human intervention and focuses on manipulation and on coordinating the steps.
This is a demonstration presented by the manufacturer, not an independently measured performance. It will be on show at Station-Berlin, hall 3, booth 3031, from 20 to 22 October. NEURA has not disclosed a price or a commercial release schedule for Neuron AI.