Using navigation

Using navigation in DimOS means exactly this: run a blueprint that includes the nav stack, give it a goal, and the stack plans a path and drives the robot there, replanning as the map updates. There are three ways to send that goal - pick the one that matches how you are building. To follow along without hardware:

dimos --simulation run unitree-go2      # robot actually moves (MuJoCo)
dimos --replay run unitree-go2          # watch the stack plan over a recorded session

On a real Go2, read the platform guide first.

Surface 1: click in the viewer (zero code)

With the stack running, open the web viewer and click a point on the map. The click is published as a clicked_point stream, and the navigation module consumes it as a goal: the costmap updates, a path appears, and the robot drives. This is the fastest way to sanity-check a nav stack, and the right first test after any mapping or planner change: if click-to-navigate works, the whole chain (map, costmap, planner, controller) works.

Surface 2: from Python (programmatic)

The entire user-facing navigation API is one small protocol, NavigationInterfaceSpec (dimos/navigation/navigation_spec.py):

class NavigationInterfaceSpec(Spec, Protocol):
    def set_goal(self, goal: PoseStamped) -> bool: ...
    def get_state(self) -> NavigationState: ...
    def is_goal_reached(self) -> bool: ...
    def cancel_goal(self) -> bool: ...

Any module in your blueprint can declare a Spec reference to the navigation module and drive it:

from dimos.core.module import Module
from dimos.navigation.navigation_spec import NavigationInterfaceSpec


class Patroller(Module):
    _navigation: NavigationInterfaceSpec

    def patrol_to(self, goal_pose) -> None:
        self._navigation.set_goal(goal_pose)
        # poll self._navigation.is_goal_reached(), or cancel with cancel_goal()

The blueprint injects the matching navigation module at build time - this is the same mechanism the built-in navigation skills use. For driving a running stack from an external script, see the Python API.

Surface 3: natural language (through the agent)

On agentic blueprints, navigation is exposed as skills: navigate_with_text("kitchen"), tag_location("kitchen"), stop_navigation(). The skill resolves language to a pose (tagged locations, then live vision, then the CLIP place map) and calls set_goal - surface 2 with a resolver in front.

dimos agent-send "go to the kitchen"                     # via the LLM
dimos mcp call navigate_with_text --arg query=kitchen    # same skill, no LLM

Full walkthrough: Tutorial: drive the Go2 with language.

Monitoring and canceling

Watch it: the viewer shows the map, costmap, and planned path live; dimos log -f shows planner state transitions.
Programmatic: get_state() and is_goal_reached() on the interface; cancel_goal() to abort.
Language: “stop” (the agent calls stop_navigation()).

Live mapping or premap?

Situation	Use
Exploring a new space	Live mapping (default `unitree-go2`) - the map builds as you drive
Returning to a known space, want a drift-free map	Premap + relocalization - record once, optimize offline, relocalize at runtime

How the map and planner actually work under the hood: How navigation works.

Getting Started

Build with DimOS

Agents

Capabilities

Platforms

Going deeper

Contribute

Surface 1: click in the viewer (zero code)

Surface 2: from Python (programmatic)

Surface 3: natural language (through the agent)

Monitoring and canceling

Live mapping or premap?

​Surface 1: click in the viewer (zero code)

​Surface 2: from Python (programmatic)

​Surface 3: natural language (through the agent)

​Monitoring and canceling

​Live mapping or premap?

Surface 1: click in the viewer (zero code)

Surface 2: from Python (programmatic)

Surface 3: natural language (through the agent)

Monitoring and canceling

Live mapping or premap?