Cloud robotics for independent researchers

Your robotics lab, without the workstation.

Describe a simulation in plain English and get a verified Isaac Sim scene on a cloud GPU — with a live viewer, full traces, and artifacts you can rerun. Pay per minute, capped by your budget.

Start with free credits See how a run works

No credit card. Idle shutdown and budget caps on by default.

Unitree G1 in a simulated yoga studio, built by a workflow agent from a plain-English prompt. — prompt> high-end yoga studio with a Unitree G1 on a mat, holding a pose

Built by a workflow agent in one run.

How it works

Describe a simulation. An agent builds it, a judge verifies it, and you keep everything.

01 — Describe

Write a plain-English prompt

Scene, robot, task, and what success looks like. No USD files, no scene graphs, no local install.

Build a Franka simulation that stacks three cubes on a flat table.

02 — Agent builds

Builder and judge loop on a cloud GPU

A coding agent assembles the Isaac Sim scene; a judge checks it against your prompt and sends it back until it passes.

session ready · builder turn 3 · judge accepted

03 — Inspect and reuse

Watch it live, keep the artifacts

Every run streams to a browser viewer and saves a versioned environment, USD scene, and full trace you can rerun.

scene.usd · env_2d4cd809 · 181 tool calls

The platform

Three ways in. One workspace.

Start with a prompt, drop into a live session when you need hands-on control, and script everything through the SDK once your experiment stabilizes.

Workflows

Prompt-to-simulation runs

The highest-level path. Describe the scene, set a turn budget, and watch the builder/judge trace with an embedded viewer.

Start a workflow →

Sessions

Hands-on Isaac Sim, on demand

A live GPU workspace with a browser viewer, SSH, and MCP access. Save any session as a reusable, versioned environment.

Browse sessions →

SDK + MCP

Programmatic everything

A Python SDK for training loops and checkpoints, plus an MCP server so agents can drive your simulations.

Read the docs →

app.cyberneticphysics.com/workflows/wfr_a9c927d2

Embedded Isaac Sim viewer showing a completed workflow scene.

Run wfr_a9c927d2

COMPLETED21m 31s

TURNS

6 / 8

TOOLS

181

JUDGE

ACCEPTED

SAVED

scene.usd

Every run keeps its trace — prompt, agent turns, judge feedback, and artifacts — so results are reproducible, not anecdotal.

Works with your tools

Your agent already knows how to use it.

Add one MCP server and Claude Code can launch simulations, inspect live sessions, and pull traces for you. The same API powers the dashboard, so anything you can click, your agent can call.

Connection guide in the docs →

~/.claude/mcp.json

{
  "mcpServers": {
    "cyberneticphysics": {
      "transport": "http",
      "url": "https://api.cyberneticphysics.com/mcp"
    }
  }
}

Pricing

Pay for GPU minutes. Nothing else.

Start with free research credits. Sessions bill per minute with idle shutdown and hard budget caps — a typical workflow run costs about a dollar.

Researcher

Free credits to start

For individual researchers. Sign in and run your first workflow in minutes.

Workflows, sessions, and viewer access
Traces, artifact storage, and versioned environments
API, SDK, and MCP access
Budget caps and idle shutdown by default

Start with free credits

Team

Shared workspaces

For labs and teams that need shared environments and controls.

Team workspaces with access controls
Shared environments and experiment history
Priority compute allocation
Per-member budgets and dedicated support

Talk to us

FAQ

Common questions

What simulation environments are supported?

Isaac Sim and Isaac Lab, plus custom URDF/USD environments. Built-in platforms include Unitree G1 and H1, Franka, UR10, ANYmal, and standard manipulators.

Do I need robotics experience to use workflows?

No. Workflows take a plain-English prompt and handle scene assembly, physics configuration, and verification. Sessions and the SDK are there when you want full control.

How do experiments stay reproducible?

Every run produces a full trace: prompt, agent turns, judge feedback, simulation parameters, and versioned artifacts. Rerun any experiment from its trace.

What does a run actually cost?

Sessions bill per GPU minute — recent RTX 4090 workflow runs on the platform cost between $0.05 and $1.60. Idle shutdown and budget caps are on by default.

Run your first simulation before your coffee cools.

Free research credits, no install, and a verified Isaac Sim scene from a one-sentence prompt.

Start with free credits Read the docs