EMOS
Cognition Recipes Overview#
Build intelligent agents from the ground up using EmbodiedAgents, the EMOS intelligence framework. These recipes introduce the core Components – the modular building blocks that drive your physical agents.
Every capability – hearing, speaking, seeing, thinking – is a component you wire together in pure Python. No ROS XML, no boilerplate.
Your first agent – wire STT, VLM, and TTS into a multimodal dialogue system.
Run LLMs, VLMs, STT, and TTS entirely on-device — no server required.
Shape agent behavior with dynamic Jinja2 templates at the topic or component level.
Give your robot a graph-backed spatio-temporal memory built on eMEM – indexed simultaneously by meaning, location, and time.
Turn “go to the kitchen” into a PoseStamped goal – an LLM with Memory tools, plus a regex preprocessor.
Generalise the tool calling pattern: write your own Python function as the LLM’s tool, with full control over the schema and the published output.
Typed questions answered in one forward pass: is this meant for me, which room is this, is the task done. Probabilities an event can act on, from a model that generates no text.
Route messages to different graph branches based on meaning, not topic names.
Combine perception, memory, and reasoning into a fully embodied system.