Blog
NEURA Robotics introduces Neuron AI, a robot model that combines vision, touch, sound and memory
On 14 October 2026 NEURA Robotics introduced Neuron AI, a multimodal Physical AI foundation model that combines 3D vision, touch, audio, robot state and language with memory. NEURA says it runs across five of its robot types. A humanoid laundry demo is planned for NVIDIA GTC Berlin on 20–22 October.
On 14 October 2026 German robotics company NEURA Robotics introduced Neuron AI, which it calls its multimodal foundation model for "Physical AI". The company says the model lets robots act on information that cameras alone cannot capture, and remember objects after they leave view.
What NEURA says the model does
NEURA calls the design a Perception-Language-Action (PLA) architecture. Rather than treating each sensor separately, the company says Neuron AI encodes the following into one shared space over time:
- 3D vision
- touch and force data
- audio
- robot state (the robot's own position and movement)
- language
It then turns those inputs into a single stream of actions.
NEURA's examples show why senses other than sight matter. When a robot opens a lock, sound and touch confirm that it has actually opened. Touch also lets a robot grip objects it cannot see. In welding, the sound of the process can show the quality of the seam.
The model also has memory. NEURA says a robot has to track objects and the results of its own actions over time, often without seeing them, much as people learn object permanence. The company says it changed the architecture to add memory and trains the model on datasets built to use it.
How it is trained
NEURA describes three data sources:
- Real-world robot data from its "NEURA Gyms", where robots of different types do tasks and record synchronised vision, touch, force, audio and movement data.
- Simulation-amplified data, generated by NEURA's Data Amplifier and simulation stack from variations of tasks and environments.
- Human demonstrations recorded by people wearing sensor-equipped suits.
NEURA says one shared model runs across five NEURA robot embodiments, each with different grippers and hands. Industrial partners including Schaeffler and Bosch supply data. The company says the model was developed and trained in-house in Metzingen, Germany.
From one skill to a fleet
Customers can fine-tune Neuron AI for their own tasks and roll out the resulting skills across robot fleets on NEURA's Neuraverse platform. NEURA says customers can keep their skills private. Any skills or deployment data they choose to share may feed future model development. The company's main measure of progress is "Time-to-Skill", meaning how fast a robot can learn a new task and do it reliably.
The demo
At NVIDIA GTC Berlin 2026 (Station-Berlin, Hall 3, Booth #3031, 20–22 October), NEURA plans to show its humanoid 4NE1 running an eleven-minute laundry sequence with no human intervention. The robot will sort laundry, load a washing machine, close and reopen it, unload it, and move the laundry to a dryer. NEURA says the demo is about handling and coordinating tasks within the laundry process.
What we don't know yet
The announcement gives no parameter count, no benchmark results, no pricing and no general availability date. It does not say whether outside researchers can access the model. Every capability claim above is NEURA's own, and none has been tested independently. The live demo has not happened yet, so we have nothing to report on how it goes.
Why it matters
Most robot foundation models discussed publicly lean heavily on vision and language. NEURA is betting that touch, sound and memory are what will make robots dependable on factory and household tasks. For European industry, the in-house German training and data from manufacturers like Bosch and Schaeffler are also part of the pitch. Whether the approach actually shortens Time-to-Skill will only become clear once customers use it.