laser-amcl.md

layout	title	package	category	namespace	tags
reference	laser_amcl	amcl	node	/laser_amcl	amcl localization particle filter

Description

Mayfield's fork of amcl is almost identical to the upstream ROS-indigo package but has a few custom changes that we made for Kuri. See the normal amcl docs for the majority of the documentation for amcl. Below, only the custom changes we made for Kuri are documented (topic names are those used on Kuri).

The major changes we made are as follows:

• Initialize amcl with multiple hypotheses
• Leave behind particles to account for getting stuck
• Localization start and stop services
• Penalize particles that end up deep in unknown space
• Visualization topics for doing localization comparisons

Subscribed Topics

/initialpose_cloud

HypothesisSet.msg
Used to seed the particle cloud with multiple hypotheses with covariances.

/initialpose
geometry_msgs/PoseWithCovarianceStamped
Normal amcl particle cloud initialization (single Gaussian).

/scan
sensor_msgs/LaserScan
The incoming laser scan.

/tf
tf/tfMessage TF transforms.

/tf_static
tf2_msgs/TFMessage
Static TF transforms.

/map
nav_msgs/OccupancyGrid
The map to use.

Published Topics

/amcl_adjusted_scans

visualization_msgs/Marker
If draw_laser_points is true, laser scans transformed to the current estimated pose will be published on this topic.

/laser_amcl_basic_pose

geometry_msgs/PoseStamped
If publish_basic_pose is true, the estimated pose is broadcast on this topic, for easy visualization in RViz.

/laser_amcl_pose

geometry_msgs/PoseWithCovarianceStamped
amcl's normal pose topic.

/laser_particlecloud

geometry_msgs/PoseArray
amcl's normal particle cloud topic. If draw_weight_as_height is true, the z-value will be proportional to the particle weight (max at 1 m).

/tf
If publish_test_frame is true, will send the current pose estimate (test_frame_id to map transform) as a broadcasted TF transform.

Services

/localization_start

std_srvs/Empty
Used to start amcl (if auto_start is true, don't need to call this to start it up)

/localization_stop

std_srvs/Empty Used to pause amcl

/set_map

nav_msgs/SetMap
Used to set the map used in amcl

Parameters

auto_start

boolean, default false
If set to true, amcl will start without you having to call the localization start service.

stuck_prob

double, default 0.0
Leave the first (current sample_count * stuck_prob) particles where they were when the action model is applied, to cover the case where the robot is actually stuck in place.

init_global

boolean, default false
If true, uses the uniform pose generator to initialize the particle filter.

draw_weight_as_height

booean, default false
If true, the height of the poses in the geometry_msgs::PoseArray broadcast on /laser_particlecloud (default topic particlecloud, on Kuri the laser_ part comes from the topic_prefix in amcl.launch) will have height (z-values) set to be proportional to the particle weight (highest particle at 1 m).

use_cov_from_params

boolean, default false
If true, uses the params std_xx, std_yy, and std_tt to initialize the particle cloud (whenever a message is received on /initialpose).

use_tf_to_update_initial_pose

boolean, default true
If false, does not attempt to integrate the odometric change that has happened in-between an /initialpose time stamp and the current time. (We have found that leaving this true can occasionally result in very crazy behavior when there is any weirdness with tf, via time jumping or running low on CPU or any such thing.)

std_xx

double, default 0.25
Only active if use_cov_from_params is true
Variance in x to use in initializing the particle cloud.
(the default value is 0.5 meters squared.)

std_yy

double, default 0.25
Only active if use_cov_from_params is true
Variance in y to use in initializing the particle cloud.
(the default value is 0.5 meters squared.)

std_tt

double, default 0.121846
Only active if use_cov_from_params is true
Variance in theta to use in initializing the particle cloud.
(the default value is 20 degrees squared.)

Params for penalizing poses that are in unknown space:

penalize_unknown

boolean, default false
If set to true, penalizes particles if a certain percentage of map pixels around them are unknown (unknown_threshold). The weighting scheme is linear. So if 60% of the pixels are unknown, we multiple the particle's existing weight by .4 (or 1-.6). This is to allow particles to somewhat travel through bits of the map with specks/shards of unknown pixels without being penalized.

unknown_radius

int, default 4
Pixel radius of the square around the robot pose that gets checked for unknown pixels (it checks from pose-square - unknown_radius to pose-square + unknown_radius in both x and y). (The default on Kuri uses 5 cm pixels, so the window is 9 * 5 = 45 cm wide.)

unknown_threshold

double, default 0.6
The percentage of map pixels around a particle's robot pose that can be unknown before a penalty is applied (if penalize_unknown is set to true).

unknown_min_penalty

double, default 0.2
Minimum penalty to apply to a particle with more than unknown_threshold percent of map pixels around them that are unknown.

Params for using beam skipping with the likelihood_field_prob laser model:
The likelihood_field_prob model doesn't use the ad-hoc weighting scheme for combining beam probs that the likelihood_field model does (taking probabilities to the third power, which seems rather arbitrary, but serves to keep around less-likely hypotheses for much longer). This mode converges much faster and generates more accurate pose estimates when the world matches the map well, since it makes full use of all laser readings to squash non-matching particles. However, it is really only useful when you have a nice laser that generates a wide field-of-view with quite accurate measurements. This mode also needs beam-skipping to avoid killing off correct particles due to dynamic obstacles.

If you have a less-accurate, lower-range, consumer-grade sensor in your robot (like Kuri's very custom depth sensor), you probably want to use the likelihood_field model, which is what we use by default on Kuri.

This stuff got pushed into normal amcl, but is not in the normal amcl documentation as of writing this.

do_beamskip

boolean, default false
Only applies if laser_model_type is set to likelihood_field_prob. If true, ignore beams that do not agree with the map for many particles' poses, to avoid down-weighting correct particles because of unexpected obstacles such as humans.

In particular, laser beams that are not within beam_skip_distance of at least (particle count * beam_skip_threshold) particles will not be used to update particle probabilities in the observation update (beam is skipped), unless the number of skipped beams is greater than (beam_skip_error_threshold_ * number of laser beams), at which point we assume the filter may have converged to the wrong solution and thus integrate all the beams after all.

beam_skip_distance

double, default 0.5
Only active if do_beamskip is true and laser_model_type is likelihood_field_prob.
If the laser beam endpoint is not within beam_skip_distance of the closest obstacle in the map, it is a candidate for beam skipping.

beam_skip_threshold

double, default 0.3
Only active if do_beamskip is true and laser_model_type is likelihood_field_prob.
If a laser beam does not agree with more than (particle count * beam_skip_threshold), it is added to the list of beams to skip.

beam_skip_error_threshold_

double, default 0.9
(yes, the trailing underscore is actually part of the param name, whoops.)
Only active if do_beamskip is true and laser_model_type is likelihood_field_prob.
If the number of skipped beams is greater than (beam_skip_error_threshold_ * number of laser beams), we assume the filter may have converged to the wrong solution and thus integrate all the beams after all.

Params for doing localization comparisons with multiple copies of amcl:

publish_basic_pose

boolean, default false
If true, will publish the current amcl pose estimate as a geometry_msgs::PoseStamped message on topic "/laser_amcl_basic_pose". (The topic name here includes the prefix "laser", which is set in amcl.launch. The original topic name is "amcl_basic_pose".)

publish_basic_pose_on_convergence

boolean, default false
If true, will only publish the current amcl pose estimate as a geometry_msgs::PoseStamped message on topic "/laser_amcl_basic_pose" when the particles are converged (all particles are within 0.5 m of the mean).

draw_laser_points

boolean, default false
Set this to true to draw the current laser scan, transformed to the current amcl pose estimate (this can be used to visualize comparisons when running multiple copies of amcl with different params). The transformed laser scan is broadcast as a visualization_msgs::Marker message on topic /amcl_adjusted_scans.

laser_colors

double[3], no default
Only active if draw_laser_points is true
If this is actually a 3-list, the values should be between 0 and 1, and correspond to the (r, g, b) values for the color that should be used to draw the laser scan.

publish_test_frame

boolean, default false
If true, sends the current amcl pose estimate to the tf broadcaster as a test_frame_id to map transform.

test_frame_id

string, default "test_map"
Only active if publish_test_frame is true
What to call the tf frame that gets sent to the tf broadcaster.

odom_only

boolean, default false
If true, updates the estimated pose only using odometry estimates

Launch File

laser_amcl.launch