Intent.
Into motion.

Give the agent a task.

The ORVYN simulator provides a direct loop between natural language planning, deterministic path controllers, and full rigid-body physics.

From intent to physics execution.

ORVYN decouples natural language reasoning from low-level joint actuators through a validated deterministic execution pipeline.

The server validates that all proposed motions map to safe, bounded controllers before sending actuator inputs to the MuJoCo simulation worker.

An ordinary world. Specific goals.

The simulation tests embodied agent competence across navigation, search, and dynamic object manipulation inside an authentic residential layout.

MARS mobile manipulator specs.

MARS combines a stable differential mobile base with an articulated inspection manipulator and binocular stereo head.

Ground-truth vs sensor feed.

In ORVYN, the robot perceives its environment through two distinct modalities:

Review intent before motion.

The backend translates natural language prompts into bounded robotic intent via the configured Orbio chat model.

The model can select strictly from four authorized symbolic skills: explore, rescue, shepherd, or stop. An AI plan is a hypothesis, not a guarantee of task completion: the physics simulation must physically execute the trajectory and satisfy upstream spatial judges.

{
  "intent": "Find the person and stay nearby",
  "plan": {
    "skill": "rescue",
    "summary": "Navigate toward human coordinates and maintain <1.2m proximity",
    "parameters": { "target": "human", "tolerance_meters": 1.2 }
  },
  "reviewed": false
}

The human-in-the-loop interface requires selecting Run reviewed plan after inspecting the model’s intent summary, ensuring transparent agent behavior.

Security and rate limiting.

ORVYN enforces strict server-side boundaries to protect compute resources and model API allocations:

• API Rate Limit: Requests are throttled to one every 15 seconds, and capped at 20 per hour across active sessions.
• State Invalidation: Triggering manual driving, pressing Stop, clicking Reset, or closing the playground immediately invalidates any pending or reviewed AI plan.
• Zero Client Storage: Credentials and model keys remain exclusively in server process memory. The browser never receives or stores secret keys.
• Deterministic Fallback: Built-in skill controllers and manual controls operate independently of AI availability, requiring zero provider balance.

You keep the controls.

Take manual control over the MARS mobile base at any moment using keyboard keys, touch buttons, or the virtual directional pad.

Keyboard driving activates when either the Camera stage or the Keyboard Driving pad is focused. Keyboard events inside inputs or menus do not move the robot.

Fail-safe mechanical watchdogs.

Robotic control requires active liveness verification. ORVYN implements a multi-layer watchdog:

Running on private infrastructure.

The ORVYN deployment runs securely on server loopback ports, isolated from unauthorized external access. The web server listens on port 8092 and forwards validated API requests to the physics daemon on port 8091.

ssh -N -L 8091:127.0.0.1:8092 web3
# Open http://127.0.0.1:8091 in your browser

Provisioning AI sessions securely.

For an approved AI planning session, invoke the secure remote starter script from the ORVYN root directory:

sh scripts/start-secure-remote.sh
# Enter the approved model ID and API key at the masked terminal prompt

Know what you are seeing.

The landing page features editorial concept visualizations. Inside the simulator, the feed is generated in real time by the actual MuJoCo physics camera, rendered at calibrated resolution and refresh rate.

This is simulated embodied AI research, not physical-robot deployment. No financial tokens, wallets, or on-chain mechanisms are required or active.