GraphNav Initialization

When a map is initially loaded, GraphNav is unaware of the robot’s location relative to it. The GraphNav service provides a set of basic primitives to initialize the robot’s location to the map. Each one requires various levels of help from the user.

• Using fiducials to initialize the robot’s localization

• Using local geometry and texture around the robot to initialize the robot’s location

Initializing with fiducials

This is the easiest method of initializing the robot’s pose to the map. In the Spot SDK, the different types of fiducial initialization are represented with the FiducialInit enum:

• FIDUCIAL_INIT_NEAREST: Localize to the nearest fiducial in any waypoint.

• FIDUCIAL_INIT_NEAREST_AT_TARGET Localize to nearest fiducial at the target waypoint.

• FIDUCIAL_INIT_SPECIFIC: Localize to the given fiducial at the target waypoint.

Fiducials are required for initialization in certain cases:

• At the beginning of every mission (Autowalk only)

• In “feature desert” areas where there is not enough information for Spot’s sensors to determine its location in the environment, fiducials might help prevent the robot getting lost

About fiducials

Fiducials are specially designed images similar to QR codes that Spot uses to match its internal map to the world around it.

Spot recognizes AprilTag fiducials that meet the following requirements:

• Fiducials that are part of the AprilTag Tag36h11 set

• Image size: 146mm square

• Printed on white non-glossy US-letter size sheets

Spot is configured to treat fiducials as 146mm squares and computes the distance based on that size. If a fiducial is printed at another size, Spot will infer the wrong position.

Boston Dynamics provides a small sample of correctly sized AprilTag fiducials that are recognized by the Spot robot in Spot Autowalk: Using AprilTag Fiducials, part of the Spot user documentation. Printing this set at 100% scaling will produce the correctly sized fiducial images.

Background information about AprilTags can be found at the University of Michigan April Robotics Lab website.

Placing fiducials

When placing fiducials, be sure to:

• Place one at the mission’s starting location.

• Tape fiducials flat against a vertical wall, as securely as possible. If fiducials move after recording, the mission may no longer replay.

• Place them low on the wall. The top of the fiducial image should be at human knee height, 45-60cm (18-24”) above the ground.

• Place them in a permanent location where they will persist for as long as you plan to replay the mission.

• Place them in areas which are feature deserts, for example a span of over 3m of featureless white wall.

When placing fiducials, avoid the following:

• Repeating the same fiducial multiple times in a single mission. Each fiducial in a mission must be unique, however the same fiducial can be used in multiple missions.

• Placing fiducials in areas with inconsistent lighting. Shadowed or unevenly lit fiducials can have unreliable detections.

• Placing fiducials so that they are backlit by a bright background, such as on a window.

• Placing fiducials where they will be blocked, damaged, moved, or removed.

Note that areas with intersecting walls, corners, furniture, equipment, and other visually distinguishing features do not typically require a fiducial (unless the mission starts in that location). Using a LIDAR payload for navigation will reduce the number of fiducials required.

Initializing with search

If fiducials are not available, the client program can use other methods of initialization through the SetLocalization RPC (available in the graph_nav.proto). The client can provide an initial guess for the complete localization in the initial_guess field which describes the robot’s pose relative to a known waypoint and which waypoint to initialize to. Details on the algorithm that is run when an initial guess is provided are described in the next section.

If the initial guess for the localization is unknown, then GraphNav can perform a brute force search. The parameters of that search are set in the SetLocalization RPC via max_distance and max_yaw fields. Depending on the size of these parameters, the SetLocalization RPC can take a long time to complete.

Initializing with local geometry and texture

Spot can run a proprietary initialization algorithm that compares local geometry and texture from where the robot is now, to similar data collected at record time. In the API, we call this a Scan Match. This technique is best suited in areas with unique features (e.g. corners, doorways); it will work poorly in areas with non-unique features or no features (e.g. bland hallways, big open spaces).

For your reference, the Spot base platform considers anything < 3 meters away “local geometry.” Spot with Lidar considers anything < 100 meters away “local geometry.”

The client must provide a reasonable guess as to where the robot is relative to a known waypoint on the map. If that guess is reasonably close, and the features around the robot are rich enough, scan match initialization can establish initial map pose without needing fiducials.

This is accomplished by leaving max_distance and max_yaw unset in the SetLocalization and by specifying FIDUCIAL_INIT_NO_FIDUCIAL for the FiducialInit enum.