[Software] Open-source Embodied Operating System for Physical AI

[Maria Kabtoul]

Dear all,

We are excited to present EMOS: the Embodied Operating System, an open-source unified orchestration layer for Physical AI.

EMOS lets you build robot agents that perceive, reason, and act. Write applications once in Python as a Recipe, and deploy across quadrupeds, humanoids, and wheeled robots without modification. It runs on top of ROS2 and is MIT-licensed.

A Recipe composes from a library of ready-to-use components, including:

• Autonomous navigation: planning, control, and drive management, fully GPU-accelerated (GPGPU via SYCL, cross-vendor).
  
• Perception: vision, depth, object detection, mapping, target tracking.
  
• Cognition: LLMs, VLMs, and VLAs for multi-modal reasoning and manipulation.
  
• Interaction: speech, dialogue, tool calling.
  
• Spatio-temporal memory: the first memory system for embodied agents built on neuroscience principles. A hybrid graph indexed by meaning, location, and time. Episodes, entities, and interoception in one store.
  
• Agentic planning: drop one component into a recipe; it auto-discovers every other capability, decomposes high-level goals, dispatches steps, watches feedback, and replans on failure. An agentic harness for embodied intelligence.
  
• Event-driven & robustness primitives: declarative events and actions; algorithm fallbacks, recovery maneuvers, model swaps. Failures are first-class control states, not crashes.
  

What you can do with EMOS:

• Build AI-native behavior. Foundation models are first-class citizens. LLMs, VLMs, and VLAs share the same lifecycle as your controllers. Object detection can trigger a controller switch; a VLM can alter planning strategy; vision components can drive target following.
  
• Run navigation on the GPU, on any GPU. Up to 3,106× speedup over CPU baselines on geometric planning, with cross-vendor SYCL support, runs on Nvidia, AMD, and Intel GPUs.
  
• Treat failures as a control state, not a crash. Cloud API drops → local model fallback. Stuck controller → recovery maneuver. Event-driven adaptive reconfiguration, all declarative.
  
• Get a web dashboard for free. EMOS auto-generates a real-time UI directly from your Recipe. Video feeds, configuration controls, no frontend code required.
  
• Plug in any model server. Ollama, vLLM, SGLang, LeRobot, RoboML, and any OpenAI-compatible endpoint.
  
• Let an LLM write your Recipes. EMOS publishes an llms.txt; feed it to your coding agent of choice and let it generate Recipes for you.
  

The documentation includes many reference Recipes covering conversational agents, semantic mapping, tool calling, VLA manipulation, point navigation, vision-guided following, runtime fallbacks, and event-driven cognition.

EMOS is developed in collaboration between Automatika Robotics and Inria. Contributions, issues, and Recipes shared back are very welcome.

Documentation: https://emos.automatikarobotics.com/

Code (monorepo): https://github.com/automatika-robotics/emos

Quickstart: https://emos.automatikarobotics.com/getting-started/quickstart.html

Discord: https://discord.gg/B9ZU6qjzND

If EMOS is useful to your work, a ⭐ on GitHub helps others discover the project.

Best Regards,

Automatika Robotics

https://automatikarobotics.com/