<div align="center">
<img src="apps/homepage/public/images/omega-logo.png" alt="OMEGA" width="180">
</div>
<h1 align="center">OMEGA</h1>
<h3 align="center">
Scalable Embodied Agent Harness for Real-World Robots
</h3>
<h4 align="center">
<a href="https://zheruiqiu.github.io/omega/">Homepage</a>
·
<a href="https://zheruiqiu.github.io/omega/docs/">Docs</a>
·
<a href="https://github.com/zheruiqiu/OMEGA/blob/main/deploy/multi-device/README.md">Multi-Device Runbook</a>
·
<a href="#architecture">Architecture</a>
</h4>
<div align="center">
<p>
<img alt="Python 3.11+" src="https://img.shields.io/badge/Python-3.11%2B-3776AB?style=for-the-badge&logo=python&logoColor=white">
<a href="LICENSE">
<img alt="License: Apache-2.0" src="https://img.shields.io/badge/License-Apache--2.0-0E8A16?style=for-the-badge">
</a>
</p>
</div>
OMEGA stands for **Orchestration System for Multiple Embodied Generalist Agents**.
Describe a mission once, and OMEGA turns it into coordinated robot work: it discovers
the available workers, plans the steps, schedules ready tasks, dispatches execution, and
keeps mission evidence attached to the run.
> [!IMPORTANT]
> **Real-robot validated** — OMEGA has been validated on a real-world robot mission.
## Why OMEGA
- **Complex robot work needs coordination.** Real-world robot missions are not single
model calls or isolated actions. OMEGA turns an operator request into a coordinated
team workflow across multiple embodied agents.
- **The architecture is built to scale.** OMEGA targets the step after a successful
robot demo: organizing isolated robot capabilities into a durable system that can
grow across agents, skills, and deployment environments.
- **Embodied AI needs a path to real-world adoption.** A robot demo proves a capability,
but real deployment depends on whether that capability can be connected to robot
execution in a repeatable way. OMEGA bridges mission-level orchestration and
downstream robot control code with a clean separation, so new robot integrations can
evolve without rewriting the whole system.
## Real-World Proof
Watch the real-robot demo on YouTube:
[OMEGA kitchen demo](https://www.youtube.com/watch?v=6ZO_N9kbn_k).
OMEGA has been validated on a real-robot kitchen mission. The proof starts from one
high-level operator request and resolves it into a scheduled task graph across two
embodied agents, including cross-agent rendezvous points where one worker's output gates
the other's next step.
<p align="center">
<img src="docs/assets/readme/omega-kitchen-task-graph.svg" alt="OMEGA kitchen task graph with cross-agent rendezvous">
</p>
- The operator submits one high-level kitchen mission.
- The orchestrator compiles a task graph and assigns atomic tasks to Agent B and
Agent C.
- Agent B and Agent C execute the assigned kitchen tasks.
- Each worker exposes public A2A readiness and semantic skills while a same-host
robot runtime performs robot-specific execution privately.
- Mission events, worker reports, and run evidence remain inspectable after the run.
## Release Scope
This release publishes:
- OMEGA orchestration source code:
- mission-level orchestrator, public A2A worker interface, and operator read model;
- shared contracts, semantic skill specifications, and capability projections;
- multi-device deployment assets, runbooks, and real-robot proof materials.
- Architecture documentation.
- OMEGA project homepage.
This release focuses on OMEGA's orchestration layer. The robot runtime integration under
`external/lerobot_lab` is planned for a later public release after compliance review,
because that subtree includes both the runtime integration used by OMEGA and broader
robot-learning framework code beyond this orchestration release.
> [!CAUTION]
> **Project in Active Development** — OMEGA is evolving rapidly. Expect breaking
> changes, restructuring, and frequent updates.
## Architecture
OMEGA keeps public coordination separate from robot-private execution. The orchestrator
discovers and dispatches work over A2A; each worker accepts assigned tasks and bridges
them into its worker-private MCP runtime.
The real-robot validation uses that same split: the control host runs the orchestrator,
while each worker host exposes an A2A worker and keeps robot execution behind a private
MCP runtime adapter.
<p align="center">
<img src="docs/assets/readme/omega-real-robot-validation.svg" alt="OMEGA real-robot validation architecture">
</p>
The sequence below abstracts the validated topology into the reusable OMEGA execution
contract.
<p align="center">
<img src="docs/assets/readme/omega-a2a-mcp-sequence.svg" alt="OMEGA A2A and MCP execution sequence">
</p>
In the sequence above, the embodied-agent lane is the public worker agent plus its
A2A-to-MCP bridge. The worker exposes the A2A contract to the orchestrator while keeping
robot execution behind the worker.
- **A2A** lets the mission team scale across workers, hosts, and scenarios.
- **MCP** adapts each worker's semantic skills to robot-specific runtime execution.
- **LangGraph** runs the mission loop: plan, schedule, dispatch, observe.
- **OpenTelemetry** records spans, events, and evidence for inspection.
Worker agents are declared by public Agent Cards and semantic skill allowlists. Runtime
adapters stay behind each worker, so OMEGA can grow the team through A2A while adapting
different robot embodiments through private MCP bindings.
## Mission Lifecycle
Every mission follows the same resolved loop:
| Stage | What happens | Public artifacts |
| --- | --- | --- |
| Discover | Probe configured workers over A2A. | Worker cards, readiness snapshots, public skill lists. |
| Plan | Build a typed task graph from the mission and scenario context. | Planner inputs, runtime rules, task graph. |
| Schedule | Resolve ready tasks against dependencies and worker bindings. | Ready-task set, worker locks, retry policy. |
| Dispatch | Hand assigned atomic tasks to worker agents. | Skill request, task attempt record. |
| Observe | Reconcile reports and publish mission evidence. | Worker report, event log, evidence outputs. |
## Validated Multi-Device Deployment
The validated deployment path uses one orchestrator host and LAN-reachable Ubuntu worker
hosts. Start with the deployment hub:
- [`deploy/multi-device/README.md`](deploy/multi-device/README.md) for the deployment
assets, validated reference environment, and runbook entry points.
- [`deploy/multi-device/runbooks/README.md`](deploy/multi-device/runbooks/README.md) for
the shared lifecycle, operator worksheet, readiness model, probe policy,
troubleshooting signals, and acceptance checklist.
## Next Steps
- Visit the public homepage: [OMEGA homepage](https://zheruiqiu.github.io/omega/).
- Read the public documentation: [OMEGA docs](https://zheruiqiu.github.io/omega/docs/).
- Review the deployment path: [Multi-device runbook](https://github.com/zheruiqiu/OMEGA/blob/main/deploy/multi-device/README.md).