Cognitive Robots Neura and Nxai Want Memory for Physical AI

Related Vendors

FAULHABER Drive Systems

Taiwan Machine Tool & Accessory Builders' Association (TMBA)

SIMTOS (Seoul International Machine Tool Show)

Neura Robotics and Nxai are partnering to collaborate on a building block for the memory of cognitive robots.

The newly formed partnership between Neura Robotics and Nxai consists of two parts: It is based on Neuron AI, Neura’s multimodal foundation model for Physical AI, which combines 3D vision, touch, audio, robot state, and language with a memory function. To achieve this, the companies are combining Neura’s Physical AI stack with XLSTM, a recurrent architecture from Nxai that can efficiently maintain and update state over long sequences. In addition, Nxai plans to make its own model, based on the XLSTM architecture, available on the Neuraverse platform. It is expected to be available in the future alongside Neura’s own foundation model, Neuron AI.

“A robot that truly helps in everyday life and in industry must be able to remember,” says David Reger, founder and CEO of Neura Robotics. “With Neuron AI, we have the foundation. With the Neuraverse, we’re creating the platform on which other AI approaches can also find their way into robots. Together with Sepp Hochreiter and Nxai, we’re expanding that memory.”

Who's Bringing What?

• Neura is introducing the Physical AI Stack: its own multimodal foundation models, robotics, sensor technology, real-world training data from Neura Gyms, and the platform for industrial use.
• Nxai brings decades of AI research and a fundamentally different approach to recurrent intelligence and memory.

Why Robots Need a Memory

“Memory and experience are essential for robots to reliably solve multi-step tasks,” says Prof. Dr. Sepp Hochreiter, founder of Nxai. “XLSTM is designed for this purpose; it is fast, memory-efficient, and can maintain its state over long sequences.”

A robot operates in a loop of perception, decision-making, and action—many times per second and, under certain circumstances, for an entire shift. To do this, it needs memory on two levels. Working memory determines what is happening right now: Has the component already been inserted? Is the door open or closed? Has the start button already been pressed? Long-term, episodic memory goes a step further: A robot that has repeatedly performed a process should be able to draw on past experiences when faced with the same situation.

A Shared Platform Offers More Choices

The planned integration is intended to enable companies to tailor the Nxai model to their own tasks using the Neuraverse toolchain, test it in the Neura Gyms, and deploy it on suitable robots. This will allow customers to use different model approaches within the same training and application environment.

Neuron AI remains Neura's proprietary multimodal foundation model and will continue to be developed independently. The partnership brings an additional model option to the platform and includes joint research. Companies retain control over their process knowledge and decide for themselves whether to share data and capabilities.