REA

A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control

English · 简体中文

🚧 Continuously updated. REA is under active development. Support for more benchmark platforms (such as RoboDojo) and more real robots is added on an ongoing basis; see Platforms and the changelog. Contributions of new platforms and robots are welcome.

REA is a framework for embodied agents that route among fast policies instead of putting a large model in the control loop. It has three layers:

• System 0, capabilities that act at control rate: frozen code programs, VLA policies, LLM tool-calling controllers, motion and whole-body controllers.
• System 1, a cheap router: route memory picks the capability that did best on development episodes, a judge gate checks completion only when a capability returns, and an escalation ladder decides what runs next under one shared budget.
• System 2, the brain: it writes or selects capabilities offline, and online it only judges or advises at boundaries. No running capability waits for it.

The same interfaces connect benchmark platforms (LIBERO / LIBERO-Pro, BEHAVIOR-1K, RoboTwin, RoboDojo, ARC-AGI-3, Gymnasium), real robots (ALOHA, LeRobot SO-100/101 and Koch, OpenRAL and RPent robots, Unitree G1, RoboDojo-RealEval's dual-arm robots, your own driver) and methods (LLM agents, VLAs, code-as-policy, controllers). REA can drive a platform's environment, or be called as a policy by a platform that runs its own loop. Every episode records its decisions, actions, timing and usage in the same format.

Start in one minute

Install Node.js 22.19 or newer:

git clone https://github.com/InfRec-Cardinal/REA.git
cd REA
npm run demo            # interface demo: a grid and a two-joint fixture, no model or hardware
npm run demo:routed     # a routed episode: route memory, boundary judge, escalation ladder, sealed score
npm run demo:async      # System 1 routing while System 2 thinks in the background
npm run rea -- list     # environments, policies and presets

These demos are deterministic and need no Python, credentials or hardware; they show the interfaces, not robot performance.

Configure once, choose a backend

A configuration names the environment, the policy (or the systems), the episodes and the budget:

{
  "version": 1,
  "environment": { "type": "gymnasium", "options": { "environmentId": "CartPole-v1" } },
  "policy": { "type": "chat" },
  "episodes": [{ "id": "cartpole-0", "seed": 0, "instruction": "Balance the pole for as long as possible." }],
  "budget": { "maxSteps": 200, "maxDurationMs": 900000, "decisionTimeoutMs": 60000 }
}

Model-backed policies use any OpenAI-compatible Chat Completions endpoint. Put the provider in .env (see env.example):

REA_MODEL=your-model
REA_BASE_URL=http://localhost:8000/v1
REA_API_KEY=your-api-key

npm run rea -- init gymnasium            # writes rea.config.json
python3 -m venv runtime/gym-venv && runtime/gym-venv/bin/python -m pip install -r requirements-gym.txt
export GYM_PYTHON="$PWD/runtime/gym-venv/bin/python"
npm run rea -- doctor rea.config.json    # checks configuration and local dependencies; connects to nothing
npm run rea -- run rea.config.json --steps 10 --output outputs/first-run

Preset	What it connects	Guide
demo	Synthetic grid and joint fixtures	none needed
gymnasium	Registered Gymnasium tasks and their native action spaces	framework reference
libero	LIBERO / LIBERO-Pro suites	LIBERO
arc-agi3	Official ARC-AGI-3 SDK, games and scorecards	ARC-AGI-3
aloha	ACT dual-arm simulation; Classic ROS1 ALOHA hardware	robots, ALOHA hardware
lerobot	SO-100/101, bimanual SO followers, Koch, Robot-compatible factories	robots
openral	Manifest-driven arms, dual arms, mobile platforms and humanoids	OpenRAL and RPent
rpent	LIBERO, RoboCasa and RoboTwin environment services	OpenRAL and RPent
g1	Unitree G1 camera, safety gate and executor	Unitree G1

BEHAVIOR-1K, RoboTwin and RoboDojo have dedicated runners or adapters: BEHAVIOR, RoboTwin, RoboDojo; every platform is listed under Platforms. Hardware presets start in preview (execute: false); task success on hardware needs an evaluator, not a command acknowledgement.

Routed and asynchronous systems

systems: {
  mode: "routed",
  system0: [
    { id: "fix", kind: "program", policy: { type: "module", options: { module: "./fix-program.mjs" } } },
    { id: "vla", kind: "vla", policy: { type: "lingbot-vla", options: { url: "ws://127.0.0.1:8006" } } },
  ],
  system1: { type: "route-memory", options: { table: "./route-table.json", recoverOn: "judge" } },
  system2: { type: "judge", options: { model: "your-vision-model", threshold: 0.6 } },
  maxStepsPerRun: 50,
}

tools/route-table.mjs builds the route table from one development evaluation per capability. See routed systems and, for background thinking with per-action routing, asynchronous systems.

Platforms

New platforms and robots are added continuously; the changelog lists each addition. Supported: the adapter is included and tested against the platform's interface. New: the adapter is included and tested offline, and validation on the platform is in progress.

Benchmark platform	How REA connects	Status
LIBERO / LIBERO-Pro	libero environment; resumable frozen-program backend	Supported
BEHAVIOR-1K (R1Pro)	program backend; online-agent episode runner	Supported
RoboTwin 2.0	VLA / code candidates with a success router; RPent and OpenRAL scenes	Supported
ARC-AGI-3	official SDK, games and scorecards	Supported
Gymnasium	any registered environment and its native action space	Supported
RoboDojo (simulation and RealEval)	REA as an XPolicyLab policy through rea serve	New
Other XPolicyLab platforms (RoboTwin, RMBench)	the same XPolicyLab adapter	New
Your platform	an environment (REA runs the loop) or rea serve (the platform runs it)

Real robot	How REA connects	Status
ALOHA: ACT simulation, Classic ROS1 hardware	aloha	Supported
LeRobot: SO-100/101, bimanual SO, Koch, Robot-compatible factories	lerobot	Supported
OpenRAL robots: Franka, UR5e/UR10e, Sawyer, WidowX, OpenArm, humanoids and more	openral	Supported
Unitree G1 (service, safety gate, cerebellum)	g1	Supported
RoboDojo-RealEval: ARX X5, Piper, Piper X	the XPolicyLab adapter	New
Your robot	a Python driver class (python-robot), a JavaScript driver, or rea serve from your robot stack

Plug in your own

To add	Implement	Guide
An LLM agent, VLA, code policy or controller	decide(input) → {action}	System 0 policies
A router or a judge / brain	route(input), judge(input) or think(input)	Routers and brains
A benchmark or simulator	reset, step (and score for routed episodes)	Environments
A program-executing simulator	the JSON-lines program protocol	Program backends
A real robot	a Python driver class (python-robot) or a JavaScript driver	Real robots
A platform that runs its own loop (XPolicyLab / RoboDojo, challenge evaluators, robot stacks)	rea serve plus a small adapter; a Python client is included	Policy server

Load any of them from a configuration with type: "module", or register them on createRegistry(); nothing in src/ needs to change.

import { createEmbodiedAgent } from "rea-agent";   // inside this repository: ./src/framework/index.mjs

const agent = await createEmbodiedAgent("./rea.config.json");
try {
  const episode = await agent.run({ seed: 3 });
  const evaluation = await agent.evaluate({ outputDir: "outputs/eval" });
} finally {
  await agent.close();
}

Documentation

• Architecture: the three systems, the two runtimes, contracts and repository layout
• Framework reference: configuration, CLI, environment and policy contracts, outputs
• Routed systems · Asynchronous systems
• Extending: policies · routers and brains · environments · program backends · real robots · policy server
• Benchmarks: LIBERO · BEHAVIOR · RoboTwin · RoboDojo · ARC-AGI-3
• Robots: ALOHA, LeRobot · ALOHA hardware · OpenRAL and RPent · Unitree G1 · G1 cerebellum · LingBot-VLA
• Testing · Contributing · Changelog · Third-party notices

What is not included

Model weights, VLA servers, simulators and their assets (LIBERO, BEHAVIOR-1K / OmniGibson, RoboTwin), ASPIRE, datasets, robot calibration and recorded runs are installed separately; each guide says what an integration expects. Benchmark task prompts and API docstrings from third-party projects are not redistributed.

Citation

@article{zhou2026rea,
  title   = {{REA}: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control},
  author  = {Zhou, Pengfei and Ma, Yixing and Zhang, Zhengfeiyang and Chen, Haoyuan and Wan, Zhenglin and Wang, Hexin and Zhao, Wangbo and You, Yang},
  journal = {arXiv preprint arXiv:2610.42053},
  year    = {2026}
}

License

MIT; see LICENSE. Third-party components and attributions are listed in NOTICE.md.