Someone just open-sourced an operating system that lets you tell robots what to do in plain English, no ROS required.
The Summary
- Dimensional (dimOS) is a new open-source OS that lets developers control physical robots using natural language and Python, eliminating the need for ROS (Robot Operating System)
- Works across humanoids, quadrupeds, drones, and robotic arms with native agent integration for perception, navigation, and spatial memory
- Currently in pre-release beta with stable support for Unitree quadrupeds and experimental support for humanoids and drones
The Signal
The robotics stack has been a nightmare for a decade. You want to make a robot do something useful, you learn ROS, C++, you wire together nodes and topics and transforms until you're debugging coordinate frames at 2am. The barrier to entry is enormous. Dimensional is taking a different bet: what if the robotics OS was designed for the agent era from day one.
This isn't just a wrapper around existing tooling. Dimensional runs agents as native modules that subscribe directly to hardware streams. Your Claude agent can read from a lidar feed, query spatial memory, and send commands to motor controllers without leaving the agent context. The phrase "vibecode your robots" sounds like marketing until you realize what they mean: you describe what you want in natural language, and the system compiles that intent into motor commands.
The hardware compatibility matrix tells the real story. Stable support for Unitree Go2 quadrupeds. Beta for Unitree G1 humanoids. Experimental for MAVLink drones and force-torque sensors. They're building the Android of robotics, and they're starting where the manufacturing has already standardized.
"Build physical applications entirely in Python that run on any humanoid, quadruped, or drone."
Compare this to the current state: most robotics companies build proprietary stacks for their hardware. Boston Dynamics has theirs. Tesla has Optimus running on custom everything. Unitree ships robots but expects you to figure out the software. Dimensional is offering a unified interface layer. One install script, one Python API, multi-platform deployment.
The perception and spatial memory stack is where this gets interesting:
- Object detection and 3D projection from camera feeds
- Vision-language models for scene understanding
- Spatio-temporal RAG (retrieval-augmented generation) for memory
- Object permanence tracking across time and movement
That last one matters more than it sounds. If a robot sees a cup on a table, then turns away, then turns back and the cup is gone, it needs to know the cup existed and is now missing. That's not just computer vision. That's memory architecture designed for physical space.
The agent-native design is the unlock. Right now, if you want an LLM to control a robot, you build an API layer between the language model and the robot's control system. Every command goes through translation: natural language to structured command to robot action. Dimensional collapses that stack. Agents run inside the OS, with direct hardware access. The latency drops. The context window includes sensor feeds.
The Implication
Watch who forks this repo in the next six months. If major hardware manufacturers or their third-party developer ecosystems start building on Dimensional, it becomes the de facto standard before anyone realizes there was a standards war. The question isn't whether agents will control physical robots. The question is what OS they'll run on. This is an early answer, and it's open source.