Documentation

Docs

Build, benchmark, and ship in 30 seconds. Base URL —.

Quickstart

Install the SDK, log in, and run a benchmark across 1,024 randomized worlds.

$ pnpm add @xoled-protocol/sdk

$ xoled login

$ xoled sim run warehouse-pick-v7 --worlds 1024 --policy pi0-ft

import { Sim, deploy } from "@xoled-protocol/sdk";

// 1. describe a scene + drop a robot
const run = await Sim.run({
  scene: "warehouse-pick-v7",
  embodiment: "quadruped",
  policy: "pi0-ft",
  worlds: 1024,            // domain-randomized
});

console.log(run.sim2real);  // 0.063  (target <= 0.05)

// 2. ship the same policy to real hardware
await deploy(run.policy, { target: "ros2" });

Concepts

The network has three surfaces. SimArena compiles a scene description into parallel domain-randomized environments. Fabric schedules those environments onto operator GPUs by live price. The runtime runs the same vision-language-action policy in sim and on real hardware. Everything is metered in GPU-seconds and settles on Robinhood Chain.

Scenes & embodiments

A scene pairs an environment (warehouse-pick-v7) with an embodiment (quadruped, 7-DoF arm, humanoid) and a policy. Domain randomization varies friction, mass, lighting, and geometry per world so the policy generalizes.

Benchmarking

Each run reports the sim2real Δ — the gap between simulated and reference real performance. Lower is better; the convergence target is ≤ 0.05. Runs are reproducible from their seed.

Metering

Compute is billed per GPU-second at the live spot price. There is no seat fee and no minimum. Charges and operator rewards aggregate and settle on Robinhood Chain.

Run a node

Any machine with an RTX 4090 or better can join the pool. Install the node client, stake nothing, and serve jobs; rewards accrue per GPU-second and are weighted by uptime.

Deploy targets

The runtime ships to ROS 2, MuJoCo, Isaac Sim, WebRTC teleop, or bare metal via liboptr. The policy artifact is identical across targets — one binary, sim to silicon.

API reference

REST + the typed SDK. Authenticate with a scoped key (sim:run, compute:read, runtime:deploy). Base URL —. Manage keys on the SDK page.

Glossary

VLA: Vision-Language-Action model — takes an image + a language instruction and outputs low-level robot actions (RT-2, OpenVLA, π0, NORA).
Sim2real gap: The mismatch between a policy's performance in simulation and on real hardware; the most-cited blocker in robotics.
Domain randomization: Varying textures, lighting, friction and geometry across simulated worlds so a policy generalizes instead of overfitting.
GPU-second: The unit of metered compute. Spot-priced; the basis of node rewards and settlement.
Rollout: One simulated episode of a policy acting in an environment from start to terminal state.
Embodiment: The robot body a policy controls — quadruped, 7-DoF arm, humanoid, mobile base.
Proprioception: A robot's sense of its own joint positions and velocities, fed into the policy alongside vision.
End-effector: The business end of a manipulator — gripper, hand or tool.
Teleoperation: Human remote control of a robot, often used to collect demonstration data.
ROS 2: Robot Operating System 2 — the de-facto middleware the runtime is compatible with.