What is Neural Cube?
Neural Cube is a fully autonomous robot orchestrated by multi-agent LLMs. It's designed to interact naturally, reason transparently, and grow alongside people. It is built less as another device to use and more as a curious companion.
The Concept
Most robots are framed as tools, measured by what they can do for us, what they automate, what value they extract or add.
This project attempts to explore a slightly different premise: a robot that primarily exists to explore and learn rather than being optimized for specific tasks beforehand.
1 — It notices
sees a forest it hasn't walked beforerecognises Maya
Hey Neural Cube, let's explore together.
Sounds like fun, Maya! Let's explore the forest together.
2 — It decides what to do
It looks like I'm in a forest, and I know Maya loves stopping to watch birds. I'll try to look for one along the way.
What it can reach for
Move
Look
Sound
Routine
Browse
Simulate
move towards [the path]
move towards [the stream]
A helper agent turns the plan into real motor commands
3 — It keeps what mattered
Maya went quiet and ended up watching the heron, for a long while. She seemed to enjoy herself.
Its memory of today
fadingtraffic somewhere off past the treeline
recentwalked the forest down to the stream with Maya
lastingMaya stopped to watch a heron and didn't want to leave
Neural Cube decides what to keep: vivid moments stay, ordinary ones blur and fade
4 — And overnight, it becomes a little more itself, and smarter
While it rests, it reads back over what it remembers and learns from it. Its own beliefs and skills, not any instruction we gave it, shape what it becomes.
many memoriesthe heron, the stream, a dozen other afternoons
a few things it now believes"Maya is happiest outdoors, and would rather wander than arrive"
skills it develops"how to explore outdoors" — uneven ground, changing light, keeping someone in sight
Design Principles
Neural Cube strives to adhere to the following principles:
- Fully open-source. Every layer is inspectable, forkable, and improvable. No black boxes.
- Offline by default. It can operate entirely without an internet connection.
- Network and vendor independent. No required cloud account, no forced ecosystem, and no kill switch held by a third party.
- Privacy as a foundation. Reasoning, memory, and learning all happen on the robot itself.
- Built by a community. Feature development, testing and validation, security auditing, and ethics alignment happen in the open, with many eyes and shared standards.
- Self-learning. It continues to adapt to its environment and to the people around it, instead of staying frozen at its shipping configuration.
Long-Term Goals
The long-term vision is a shared, open foundation that the community can improve together. With that in mind, we are trying to:
- Stay interpretable so the decision-making process, including simulated what-ifs, is auditable by people.
- Stay open-ended so the robot can continue to explore and learn new things, rather than be optimized for specific tasks only.
- Stay aligned with ethical principles, so the self-learning system is shaped by ethical values, and not just the data it sees.