Embodied Agent OS
MVPAn agentic runtime that any robot platform can run.
An agentic operating layer that sits above heterogeneous robot hardware: perception in, task graph out, every actuation logged and reversible.
- Codename
- Robot-Brain
- Runtime
- Python · FastAPI · async orchestrator
- Interface
- Live operator dashboard over WebSocket
- Deployment
- Single containerised service
Overview
Robot hardware is getting cheap and capable faster than robot software is getting general. The result is a fleet where every arm, every rover and every manipulator carries a bespoke, brittle control stack written for that model alone. Embodied Agent OS attacks the layer above: an agentic runtime that reasons about a task, decomposes it, and drives whatever body is available.
The system splits into a perception and world-model layer, an orchestrator that turns an instruction into a plan and a plan into a sequence of guarded actions, and a hardware abstraction that makes a specific robot a driver rather than a rewrite. A live dashboard exposes the whole loop — what the robot believes, what it intends, and what it actually did.
Because it stands on Symbiosis Core, an embodied agent inherits the same identity, memory and audit guarantees as a software agent. A robot acting in someone's home is, in the end, a delegated authority problem.
Capabilities
Task decomposition
Natural-language intent becomes an explicit, inspectable task graph before a single actuator moves.
Hardware abstraction
New robot models arrive as drivers behind a stable interface — the reasoning layer never learns their quirks.
Guarded actuation
Every command passes a policy check and lands in the audit trail. Operators can watch, pause and take over.
Operator dashboard
Real-time view of world model, current plan, executed steps and failures, streamed to the browser.