> ## Documentation Index
> Fetch the complete documentation index at: https://docs.generalrobotics.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Agentic navigation

> Ask the CLI agent to plan a route, review it, and approve robot movement.

Use a CLI and robot workflow environment with the agent navigation tools.
Navigation must be installed on the robot. This guide does not claim physical inspection accuracy.

## Prepare the robot

Finish any notebook-controlled mission first. Use one controller for the robot.
Prepare the robot and its stop control as you would for a manual navigation test.
Connecting the CLI does not establish physical readiness.
Set `ANTHROPIC_API_KEY` in the shell that starts GRID, and sign in to GRID.
Do not put credentials in prompts or notebooks.

Inside the GRID terminal, run:

```text theme={null}
agent --robot YOUR_ROBOT_NAME help me plan a navigation test
```

## Ask for a destination

For example:

```text theme={null}
Use my office map. Find the nearest mapped chair and show me a route before moving.
```

If the map is not clear, identify it when the agent asks.
The agent checks navigation state, starts Nav if needed, finds candidates, and plans a route.
These read and planning operations do not require movement approval.
Object lookup depends on the labels and objects stored in the map.
A request for Kitchenette does not create a named place or teach its position.
If no suitable match exists, choose another supported target. Do not accept an unrelated match.

## Review and approve

The CLI opens the route preview before asking for movement approval.
The proposed route is dashed amber and marked **Proposed route — not approved**.
Review its destination, starting position, and full route.

Press `y`, then Enter, to approve. Press Enter or `n` to deny.
The approval applies to that exact route and robot session.
The CLI checks the map, localization, and starting position again before dispatch.
A changed or stale plan must be reviewed again.

The direct agent navigation workflow requires a planned Path reference.
It hides `move_to`, which would skip route review, and rejects `blocking=False` for supervised motion.
This restriction does not remove `blocking=False` from the Python notebook API.
Use the notebook workflow for Chapter 3's concurrent image processing.

## Stop and check the result

Press Ctrl+X to request a stop during an action.
A stop request is not proof that the robot stopped. Check the robot and answer the terminal's stop question.
If stopping cannot be confirmed, use the physical stop and involve the operator.
An unresolved action blocks further movement until stopping is confirmed.

Read the mission result. Successful navigation does not prove that a door was inspected or a label was read.
Press Ctrl+L to inspect agent events. Press Tab there to switch to the robot's mission log.
The terminal shows where inspected logs were saved.

## Current limits

* The agent uses available robot and Nav methods. A natural-language request does not add unsupported robot capabilities.
* General task verification and reliable generated-code execution remain incomplete.
* Camera inspection, precise alignment, and route continuation need their own explicit workflow and evidence.
* Navigation tools and generated Python programs have different cancellation behavior. Ending a Python process alone does not stop hardware.

For notebook images and measurements, see [Navigation telemetry](/cli/nav-telemetry).


## Related topics

- [Navigation through GRID CLI](/navigation/overview.md)
- [Navigation telemetry and notebook logging](/cli/nav-telemetry.md)
- [Navigation troubleshooting](/navigation/troubleshooting.md)
- [Configure navigation](/navigation/configuration.md)
- [Navigation data schemas](/navigation/data-types.md)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.