Most robotics stacks force you down to C++ and low-level control loops. OM1 flips that: it treats a robot like an agent runtime — sensors in, LLM reasoning in the middle, motor commands out — and you wire the whole thing in JSON5 instead of writing control code. The key idea is a hardware abstraction layer where the same agent definition drives a humanoid, a quadruped, an educational bot, or a pure simulator.
What Sets It Apart
- Pluggable model layer spanning OpenAI, Anthropic, Gemini, DeepSeek, xAI, Meta and Ollama, plus vision-language models. So what: you can swap the reasoning backend without touching robot code.
- A Go runtime that replaced an earlier Python implementation for lower latency and a smaller memory footprint. So what: the agent loop runs on the edge device itself rather than a tethered workstation.
- Middleware bridges to ROS2, Zenoh and CycloneDDS, with Prometheus and Grafana observability. So what: it drops into existing robotics infrastructure and lets you watch decisions in real time.
- Behavior lives in JSON5 config, not source. So what: changing what the robot does is editing a file, not recompiling.
Who It's For
Great fit if you are a roboticist or AI developer prototyping embodied agents across Unitree G1/Go2, TurtleBot 4, or Gazebo/Isaac Sim and want to iterate on behavior quickly. Look elsewhere if you need deterministic real-time control or production safety guarantees: it is still beta, LLM-in-the-loop reasoning adds latency and nondeterminism, and out-of-the-box hardware support covers only a handful of platforms.
