NEURA Robotics unveils Neuron AI, a robot model that adds touch, sound and memory

What did NEURA Robotics announce?

NEURA Robotics announced Neuron AI on 14 October 2026, calling it its multimodal foundation model for physical AI. Physical AI is the industry’s name for models that control machines in the real world rather than generate text or images. NEURA, founded in 2019 and headquartered in Metzingen in southern Germany, builds collaborative robot arms, mobile robots and the 4NE1 humanoid.

The company’s claim is that most robot models judge the world through cameras, and that this is not enough. Neuron AI combines 3D vision, touch, audio, the robot’s own state and language, and adds memory so the robot can keep track of objects it can no longer see. NEURA’s founder and chief executive David Reger said in the release that “an intelligent robot needs a brain that brings its senses together and remembers what happened,” and called the model “developed in Germany, built for the world.”

This is a product announcement, not a research paper. NEURA describes the architecture and training approach in broad terms, but has not published a technical report, model size, benchmark results or independent evaluations.

How is a perception-language-action model different from a VLA?

Most current robot foundation models are vision-language-action (VLA) models. They take camera images and a text instruction and output motor commands. Physical Intelligence’s models, which it recently opened to outside teams through its Pi API beta for fine-tuning robot models, are a well-known example.

NEURA’s Neuron AI product page describes its approach as perception-language-action, or PLA. In the company’s words, Neuron AI “does what a VLA does, and adds touch, sound and memory for what a robot cannot see.” Every modality is encoded, aligned in one shared space over time and turned into a single action stream. Vision is handled as point clouds and depth, so the scene is treated as 3D space rather than a flat image.

The memory part is meant to give the robot object permanence, the understanding that a part hidden behind another part still exists. NEURA says it adapted the architecture specifically to add memory and trains it on datasets designed to use it. That puts it in the same broad conversation as research on robot world models, such as Meta’s RoboJEPA latent world models, although NEURA has not shared how its memory works internally.

Why do touch, sound and memory matter to a robot?

NEURA’s argument is practical. Cameras capture geometry, colour and position, but nothing about contact, force or sound. People feel when a screw has caught and hear a connector click into place. The release gives three examples: when a robot opens a lock, only sound and touch confirm it is really open; with touch, a robot can grasp objects it cannot see; and in welding, the quality of a seam can be judged from the sound of the process.

Touch and force sensing are not new in industrial robotics, but they are usually handled by separate controllers or hand-written rules. Putting them inside the same learned model as vision and language is the part NEURA is betting on. If it works, a single policy could react to a part slipping or a latch failing without someone programming that case in advance.

Memory matters for long tasks. A robot loading a washing machine has to remember what it has already put in, whether the door is closed and what comes next, often while its cameras point elsewhere. Other labs are attacking the same problem from the hand side, for example Natura-1, a 50-billion-parameter robot hand model trained on human video.

What data trains Neuron AI, and which robots run it?

NEURA says Neuron AI learns from three sources of physical experience. The first is real-world robot data from robots performing tasks in NEURA Gyms, the company’s training facilities, and in real environments, recorded as synchronised vision, touch, force, audio and movement. The second is simulation: NEURA’s Data Amplifier and simulation stack create variations of tasks and settings. The third is human demonstration data, captured by people wearing sensor-equipped data suits.

The company also names data partnerships with industrial firms including Schaeffler and Bosch, which complement training in the Gyms. Neuron AI was developed and trained in-house in Metzingen.

The same model is trained across five NEURA robot embodiments with different grippers and hands. The product page lists them as 4NE-1 Gen 3, 4NE-1 Gen 4, a wheeled 4NE-1, and two further bimanual platforms. One model across several bodies is the same goal most robot foundation-model companies chase, because it lets data from one robot improve another.

How will customers deploy and fine-tune Neuron AI?

NEURA says Neuron AI is deployed locally and runs inference on the robot itself, with no round trip to a server. It pitches that as both a latency benefit and a data-control benefit, noting that production data staying under the customer’s control is “a key point for manufacturers in Europe.”

Deployment and updates go through the Neuraverse, NEURA’s software platform for robot fleets. The company says customers can use it to fine-tune the model for their own tasks and keep those skills private. If they choose to share skills or deployment data, those contributions can feed the wider ecosystem and future model releases.

NEURA calls its key measure “Time-to-Skill”: how quickly a robot can learn a new task and perform it reliably. It has not published Time-to-Skill numbers for Neuron AI, nor said whether the model is available to customers today, what it costs, or whether it can run on robots from other manufacturers.

What will the GTC Berlin demo show, and what is missing?

The first public showing will be at NVIDIA GTC Berlin 2026. NEURA says its 4NE1 humanoid will perform an eleven-minute laundry sequence with the current version of Neuron AI, without human intervention: sorting laundry, loading a washing machine, closing and reopening it, unloading it and moving the laundry to a dryer. Visitors can see it at Station-Berlin, Hall 3, Booth 3031, from 20 to 22 October.

NEURA itself says the demo focuses on handling and task coordination within the laundry process. That is a sensible test of memory and sequencing, but a staged booth demo is not the same as reliability across many runs, objects and rooms. The release does not say how many attempts succeed or how often the robot is reset.

What is missing so far:

• Benchmarks: no success rates, comparisons with VLA models or results on public robot benchmarks.
• Model details: no parameter count, compute requirements for on-robot inference or technical report.
• Availability: no release date, pricing or statement on open weights.
• Independent tests: no third-party evaluation of how much touch, sound and memory improve results over vision alone.